X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTarget%2FHexagon%2FHexagonInstrInfoV4.td;h=65b0f49743679e2d4055aa9921d123cf336f5e8e;hb=083f0122f0398137c45659518e9be00748d24b2d;hp=60b5e22632f14fdb5f7cf23ed6e4cf4be1a650f7;hpb=3d1d6d904346bcc3e0db5ca429556132225d435f;p=oota-llvm.git diff --git a/lib/Target/Hexagon/HexagonInstrInfoV4.td b/lib/Target/Hexagon/HexagonInstrInfoV4.td index 60b5e22632f..65b0f497436 100644 --- a/lib/Target/Hexagon/HexagonInstrInfoV4.td +++ b/lib/Target/Hexagon/HexagonInstrInfoV4.td @@ -11,6 +11,28 @@ // //===----------------------------------------------------------------------===// +def DuplexIClass0: InstDuplex < 0 >; +def DuplexIClass1: InstDuplex < 1 >; +def DuplexIClass2: InstDuplex < 2 >; +let isExtendable = 1 in { + def DuplexIClass3: InstDuplex < 3 >; + def DuplexIClass4: InstDuplex < 4 >; + def DuplexIClass5: InstDuplex < 5 >; + def DuplexIClass6: InstDuplex < 6 >; + def DuplexIClass7: InstDuplex < 7 >; +} +def DuplexIClass8: InstDuplex < 8 >; +def DuplexIClass9: InstDuplex < 9 >; +def DuplexIClassA: InstDuplex < 0xA >; +def DuplexIClassB: InstDuplex < 0xB >; +def DuplexIClassC: InstDuplex < 0xC >; +def DuplexIClassD: InstDuplex < 0xD >; +def DuplexIClassE: InstDuplex < 0xE >; +def DuplexIClassF: InstDuplex < 0xF >; + +def addrga: PatLeaf<(i32 AddrGA:$Addr)>; +def addrgp: PatLeaf<(i32 AddrGP:$Addr)>; + let hasSideEffects = 0 in class T_Immext : EXTENDERInst<(outs), (ins ImmType:$imm), @@ -35,19 +57,9 @@ def BITPOS32 : SDNodeXFormgetSExtValue(); - return XformMskToBitPosU5Imm(imm); + return XformMskToBitPosU5Imm(imm, SDLoc(N)); }]>; -// Fold (add (CONST32 tglobaladdr:$addr) ) into a global address. -def FoldGlobalAddr : ComplexPattern; - -// Fold (add (CONST32_GP tglobaladdr:$addr) ) into a global address. -def FoldGlobalAddrGP : ComplexPattern; - -def NumUsesBelowThresCONST32 : PatFrag<(ops node:$addr), - (HexagonCONST32 node:$addr), [{ - return hasNumUsesBelowThresGA(N->getOperand(0).getNode()); -}]>; // Hexagon V4 Architecture spec defines 8 instruction classes: // LD ST ALU32 XTYPE J JR MEMOP NV CR SYSTEM(system is not implemented in the @@ -119,21 +131,19 @@ class T_ALU32_3op_not MajOp, bits<3> MinOp, let AsmString = "$Rd = "#mnemonic#"($Rs, ~$Rt)"; } -let BaseOpcode = "andn_rr", CextOpcode = "andn", isCodeGenOnly = 0 in +let BaseOpcode = "andn_rr", CextOpcode = "andn" in def A4_andn : T_ALU32_3op_not<"and", 0b001, 0b100, 1>; -let BaseOpcode = "orn_rr", CextOpcode = "orn", isCodeGenOnly = 0 in +let BaseOpcode = "orn_rr", CextOpcode = "orn" in def A4_orn : T_ALU32_3op_not<"or", 0b001, 0b101, 1>; -let CextOpcode = "rcmp.eq", isCodeGenOnly = 0 in +let CextOpcode = "rcmp.eq" in def A4_rcmpeq : T_ALU32_3op<"cmp.eq", 0b011, 0b010, 0, 1>; -let CextOpcode = "!rcmp.eq", isCodeGenOnly = 0 in +let CextOpcode = "!rcmp.eq" in def A4_rcmpneq : T_ALU32_3op<"!cmp.eq", 0b011, 0b011, 0, 1>; -let isCodeGenOnly = 0 in { def C4_cmpneq : T_ALU32_3op_cmp<"!cmp.eq", 0b00, 1, 1>; def C4_cmplte : T_ALU32_3op_cmp<"!cmp.gt", 0b10, 1, 0>; def C4_cmplteu : T_ALU32_3op_cmp<"!cmp.gtu", 0b11, 1, 0>; -} // Pats for instruction selection. @@ -146,11 +156,15 @@ class CmpInReg def: T_cmp32_rr_pat, i32>; def: T_cmp32_rr_pat, i32>; +def: T_cmp32_rr_pat; +def: T_cmp32_rr_pat; + +def: T_cmp32_rr_pat, i1>; + class T_CMP_rrbh MinOp, bit IsComm> : SInst<(outs PredRegs:$Pd), (ins IntRegs:$Rs, IntRegs:$Rt), "$Pd = "#mnemonic#"($Rs, $Rt)", [], "", S_3op_tc_2early_SLOT23>, ImmRegRel { - let validSubTargets = HasV4SubT; let InputType = "reg"; let CextOpcode = mnemonic; let isCompare = 1; @@ -169,13 +183,26 @@ class T_CMP_rrbh MinOp, bit IsComm> let Inst{1-0} = Pd; } -let isCodeGenOnly = 0 in { def A4_cmpbeq : T_CMP_rrbh<"cmpb.eq", 0b110, 1>; def A4_cmpbgt : T_CMP_rrbh<"cmpb.gt", 0b010, 0>; def A4_cmpbgtu : T_CMP_rrbh<"cmpb.gtu", 0b111, 0>; def A4_cmpheq : T_CMP_rrbh<"cmph.eq", 0b011, 1>; def A4_cmphgt : T_CMP_rrbh<"cmph.gt", 0b100, 0>; def A4_cmphgtu : T_CMP_rrbh<"cmph.gtu", 0b101, 0>; + +let AddedComplexity = 100 in { + def: Pat<(i1 (seteq (and (xor (i32 IntRegs:$Rs), (i32 IntRegs:$Rt)), + 255), 0)), + (A4_cmpbeq IntRegs:$Rs, IntRegs:$Rt)>; + def: Pat<(i1 (setne (and (xor (i32 IntRegs:$Rs), (i32 IntRegs:$Rt)), + 255), 0)), + (C2_not (A4_cmpbeq IntRegs:$Rs, IntRegs:$Rt))>; + def: Pat<(i1 (seteq (and (xor (i32 IntRegs:$Rs), (i32 IntRegs:$Rt)), + 65535), 0)), + (A4_cmpheq IntRegs:$Rs, IntRegs:$Rt)>; + def: Pat<(i1 (setne (and (xor (i32 IntRegs:$Rs), (i32 IntRegs:$Rt)), + 65535), 0)), + (C2_not (A4_cmpheq IntRegs:$Rs, IntRegs:$Rt))>; } class T_CMP_ribh MajOp, bit IsHalf, bit IsComm, @@ -183,7 +210,6 @@ class T_CMP_ribh MajOp, bit IsHalf, bit IsComm, : ALU64Inst<(outs PredRegs:$Pd), (ins IntRegs:$Rs, ImmType:$Imm), "$Pd = "#mnemonic#"($Rs, #$Imm)", [], "", ALU64_tc_2early_SLOT23>, ImmRegRel { - let validSubTargets = HasV4SubT; let InputType = "imm"; let CextOpcode = mnemonic; let isCompare = 1; @@ -208,19 +234,17 @@ class T_CMP_ribh MajOp, bit IsHalf, bit IsComm, let Inst{1-0} = Pd; } -let isCodeGenOnly = 0 in { def A4_cmpbeqi : T_CMP_ribh<"cmpb.eq", 0b00, 0, 1, u8Imm, 0, 0, 8>; def A4_cmpbgti : T_CMP_ribh<"cmpb.gt", 0b01, 0, 0, s8Imm, 0, 1, 8>; def A4_cmpbgtui : T_CMP_ribh<"cmpb.gtu", 0b10, 0, 0, u7Ext, 1, 0, 7>; def A4_cmpheqi : T_CMP_ribh<"cmph.eq", 0b00, 1, 1, s8Ext, 1, 1, 8>; def A4_cmphgti : T_CMP_ribh<"cmph.gt", 0b01, 1, 0, s8Ext, 1, 1, 8>; def A4_cmphgtui : T_CMP_ribh<"cmph.gtu", 0b10, 1, 0, u7Ext, 1, 0, 7>; -} + class T_RCMP_EQ_ri : ALU32_ri<(outs IntRegs:$Rd), (ins IntRegs:$Rs, s8Ext:$s8), "$Rd = "#mnemonic#"($Rs, #$s8)", [], "", ALU32_2op_tc_1_SLOT0123>, ImmRegRel { - let validSubTargets = HasV4SubT; let InputType = "imm"; let CextOpcode = !if (IsNeg, "!rcmp.eq", "rcmp.eq"); let isExtendable = 1; @@ -243,22 +267,19 @@ class T_RCMP_EQ_ri let Inst{4-0} = Rd; } -let isCodeGenOnly = 0 in { def A4_rcmpeqi : T_RCMP_EQ_ri<"cmp.eq", 0>; def A4_rcmpneqi : T_RCMP_EQ_ri<"!cmp.eq", 1>; -} -def: Pat<(i32 (zext (i1 (seteq (i32 IntRegs:$Rs), s8ExtPred:$s8)))), - (A4_rcmpeqi IntRegs:$Rs, s8ExtPred:$s8)>; -def: Pat<(i32 (zext (i1 (setne (i32 IntRegs:$Rs), s8ExtPred:$s8)))), - (A4_rcmpneqi IntRegs:$Rs, s8ExtPred:$s8)>; +def: Pat<(i32 (zext (i1 (seteq (i32 IntRegs:$Rs), s32ImmPred:$s8)))), + (A4_rcmpeqi IntRegs:$Rs, s32ImmPred:$s8)>; +def: Pat<(i32 (zext (i1 (setne (i32 IntRegs:$Rs), s32ImmPred:$s8)))), + (A4_rcmpneqi IntRegs:$Rs, s32ImmPred:$s8)>; // Preserve the S2_tstbit_r generation def: Pat<(i32 (zext (i1 (setne (i32 (and (i32 (shl 1, (i32 IntRegs:$src2))), (i32 IntRegs:$src1))), 0)))), (C2_muxii (S2_tstbit_r IntRegs:$src1, IntRegs:$src2), 1, 0)>; - //===----------------------------------------------------------------------===// // ALU32 - //===----------------------------------------------------------------------===// @@ -286,26 +307,23 @@ class T_Combine1 MajOp, dag ins, string AsmStr> let Inst{4-0} = Rdd; } -let opExtendable = 2, isCodeGenOnly = 0 in +let opExtendable = 2 in def A4_combineri : T_Combine1<0b00, (ins IntRegs:$Rs, s8Ext:$s8), "$Rdd = combine($Rs, #$s8)">; -let opExtendable = 1, isCodeGenOnly = 0 in +let opExtendable = 1 in def A4_combineir : T_Combine1<0b01, (ins s8Ext:$s8, IntRegs:$Rs), "$Rdd = combine(#$s8, $Rs)">; -def HexagonWrapperCombineRI_V4 : - SDNode<"HexagonISD::WrapperCombineRI_V4", SDTHexagonI64I32I32>; -def HexagonWrapperCombineIR_V4 : - SDNode<"HexagonISD::WrapperCombineIR_V4", SDTHexagonI64I32I32>; - -def : Pat <(HexagonWrapperCombineRI_V4 IntRegs:$r, s8ExtPred:$i), - (A4_combineri IntRegs:$r, s8ExtPred:$i)>, - Requires<[HasV4T]>; +// The complexity of the combines involving immediates should be greater +// than the complexity of the combine with two registers. +let AddedComplexity = 50 in { +def: Pat<(HexagonCOMBINE IntRegs:$r, s32ImmPred:$i), + (A4_combineri IntRegs:$r, s32ImmPred:$i)>; -def : Pat <(HexagonWrapperCombineIR_V4 s8ExtPred:$i, IntRegs:$r), - (A4_combineir s8ExtPred:$i, IntRegs:$r)>, - Requires<[HasV4T]>; +def: Pat<(HexagonCOMBINE s32ImmPred:$i, IntRegs:$r), + (A4_combineir s32ImmPred:$i, IntRegs:$r)>; +} // A4_combineii: Set two small immediates. let hasSideEffects = 0, isExtendable = 1, opExtentBits = 6, opExtendable = 2 in @@ -323,6 +341,12 @@ def A4_combineii: ALU32Inst<(outs DoubleRegs:$Rdd), (ins s8Imm:$s8, u6Ext:$U6), let Inst{4-0} = Rdd; } +// The complexity of the combine with two immediates should be greater than +// the complexity of a combine involving a register. +let AddedComplexity = 75 in +def: Pat<(HexagonCOMBINE s8ImmPred:$s8, u32ImmPred:$u6), + (A4_combineii imm:$s8, imm:$u6)>; + //===----------------------------------------------------------------------===// // ALU32/PERM - //===----------------------------------------------------------------------===// @@ -330,24 +354,182 @@ def A4_combineii: ALU32Inst<(outs DoubleRegs:$Rdd), (ins s8Imm:$s8, u6Ext:$U6), //===----------------------------------------------------------------------===// // LD + //===----------------------------------------------------------------------===// + +def Zext64: OutPatFrag<(ops node:$Rs), + (i64 (A4_combineir 0, (i32 $Rs)))>; +def Sext64: OutPatFrag<(ops node:$Rs), + (i64 (A2_sxtw (i32 $Rs)))>; + +// Patterns to generate indexed loads with different forms of the address: +// - frameindex, +// - base + offset, +// - base (without offset). +multiclass Loadxm_pat { + def: Pat<(VT (Load AddrFI:$fi)), + (VT (ValueMod (MI AddrFI:$fi, 0)))>; + def: Pat<(VT (Load (add AddrFI:$fi, ImmPred:$Off))), + (VT (ValueMod (MI AddrFI:$fi, imm:$Off)))>; + def: Pat<(VT (Load (add IntRegs:$Rs, ImmPred:$Off))), + (VT (ValueMod (MI IntRegs:$Rs, imm:$Off)))>; + def: Pat<(VT (Load (i32 IntRegs:$Rs))), + (VT (ValueMod (MI IntRegs:$Rs, 0)))>; +} + +defm: Loadxm_pat; +defm: Loadxm_pat; +defm: Loadxm_pat; +defm: Loadxm_pat; +defm: Loadxm_pat; +defm: Loadxm_pat; +defm: Loadxm_pat; +defm: Loadxm_pat; + +// Map Rdd = anyext(Rs) -> Rdd = combine(#0, Rs). +def: Pat<(i64 (anyext (i32 IntRegs:$src1))), (Zext64 IntRegs:$src1)>; + //===----------------------------------------------------------------------===// // Template class for load instructions with Absolute set addressing mode. //===----------------------------------------------------------------------===// -let isExtended = 1, opExtendable = 2, hasSideEffects = 0, -validSubTargets = HasV4SubT, addrMode = AbsoluteSet in -class T_LD_abs_set: - LDInst2<(outs RC:$dst1, IntRegs:$dst2), - (ins u0AlwaysExt:$addr), - "$dst1 = "#mnemonic#"($dst2=##$addr)", - []>, - Requires<[HasV4T]>; +let isExtended = 1, opExtendable = 2, opExtentBits = 6, addrMode = AbsoluteSet, + hasSideEffects = 0 in +class T_LD_abs_setMajOp>: + LDInst<(outs RC:$dst1, IntRegs:$dst2), + (ins u6Ext:$addr), + "$dst1 = "#mnemonic#"($dst2 = #$addr)", + []> { + bits<7> name; + bits<5> dst1; + bits<5> dst2; + bits<6> addr; + + let IClass = 0b1001; + let Inst{27-25} = 0b101; + let Inst{24-21} = MajOp; + let Inst{13-12} = 0b01; + let Inst{4-0} = dst1; + let Inst{20-16} = dst2; + let Inst{11-8} = addr{5-2}; + let Inst{6-5} = addr{1-0}; +} + +let accessSize = ByteAccess, hasNewValue = 1 in { + def L4_loadrb_ap : T_LD_abs_set <"memb", IntRegs, 0b1000>; + def L4_loadrub_ap : T_LD_abs_set <"memub", IntRegs, 0b1001>; +} + +let accessSize = HalfWordAccess, hasNewValue = 1 in { + def L4_loadrh_ap : T_LD_abs_set <"memh", IntRegs, 0b1010>; + def L4_loadruh_ap : T_LD_abs_set <"memuh", IntRegs, 0b1011>; + def L4_loadbsw2_ap : T_LD_abs_set <"membh", IntRegs, 0b0001>; + def L4_loadbzw2_ap : T_LD_abs_set <"memubh", IntRegs, 0b0011>; +} + +let accessSize = WordAccess, hasNewValue = 1 in + def L4_loadri_ap : T_LD_abs_set <"memw", IntRegs, 0b1100>; -def LDrid_abs_set_V4 : T_LD_abs_set <"memd", DoubleRegs>; -def LDrib_abs_set_V4 : T_LD_abs_set <"memb", IntRegs>; -def LDriub_abs_set_V4 : T_LD_abs_set <"memub", IntRegs>; -def LDrih_abs_set_V4 : T_LD_abs_set <"memh", IntRegs>; -def LDriw_abs_set_V4 : T_LD_abs_set <"memw", IntRegs>; -def LDriuh_abs_set_V4 : T_LD_abs_set <"memuh", IntRegs>; +let accessSize = WordAccess in { + def L4_loadbzw4_ap : T_LD_abs_set <"memubh", DoubleRegs, 0b0101>; + def L4_loadbsw4_ap : T_LD_abs_set <"membh", DoubleRegs, 0b0111>; +} + +let accessSize = DoubleWordAccess in +def L4_loadrd_ap : T_LD_abs_set <"memd", DoubleRegs, 0b1110>; + +let accessSize = ByteAccess in + def L4_loadalignb_ap : T_LD_abs_set <"memb_fifo", DoubleRegs, 0b0100>; + +let accessSize = HalfWordAccess in +def L4_loadalignh_ap : T_LD_abs_set <"memh_fifo", DoubleRegs, 0b0010>; + +// Load - Indirect with long offset +let InputType = "imm", addrMode = BaseLongOffset, isExtended = 1, +opExtentBits = 6, opExtendable = 3 in +class T_LoadAbsReg MajOp> + : LDInst <(outs RC:$dst), (ins IntRegs:$src1, u2Imm:$src2, u6Ext:$src3), + "$dst = "#mnemonic#"($src1<<#$src2 + #$src3)", + [] >, ImmRegShl { + bits<5> dst; + bits<5> src1; + bits<2> src2; + bits<6> src3; + let CextOpcode = CextOp; + let hasNewValue = !if (!eq(!cast(RC), "DoubleRegs"), 0, 1); + + let IClass = 0b1001; + let Inst{27-25} = 0b110; + let Inst{24-21} = MajOp; + let Inst{20-16} = src1; + let Inst{13} = src2{1}; + let Inst{12} = 0b1; + let Inst{11-8} = src3{5-2}; + let Inst{7} = src2{0}; + let Inst{6-5} = src3{1-0}; + let Inst{4-0} = dst; + } + +let accessSize = ByteAccess in { + def L4_loadrb_ur : T_LoadAbsReg<"memb", "LDrib", IntRegs, 0b1000>; + def L4_loadrub_ur : T_LoadAbsReg<"memub", "LDriub", IntRegs, 0b1001>; + def L4_loadalignb_ur : T_LoadAbsReg<"memb_fifo", "LDrib_fifo", + DoubleRegs, 0b0100>; +} + +let accessSize = HalfWordAccess in { + def L4_loadrh_ur : T_LoadAbsReg<"memh", "LDrih", IntRegs, 0b1010>; + def L4_loadruh_ur : T_LoadAbsReg<"memuh", "LDriuh", IntRegs, 0b1011>; + def L4_loadbsw2_ur : T_LoadAbsReg<"membh", "LDribh2", IntRegs, 0b0001>; + def L4_loadbzw2_ur : T_LoadAbsReg<"memubh", "LDriubh2", IntRegs, 0b0011>; + def L4_loadalignh_ur : T_LoadAbsReg<"memh_fifo", "LDrih_fifo", + DoubleRegs, 0b0010>; +} + +let accessSize = WordAccess in { + def L4_loadri_ur : T_LoadAbsReg<"memw", "LDriw", IntRegs, 0b1100>; + def L4_loadbsw4_ur : T_LoadAbsReg<"membh", "LDribh4", DoubleRegs, 0b0111>; + def L4_loadbzw4_ur : T_LoadAbsReg<"memubh", "LDriubh4", DoubleRegs, 0b0101>; +} + +let accessSize = DoubleWordAccess in +def L4_loadrd_ur : T_LoadAbsReg<"memd", "LDrid", DoubleRegs, 0b1110>; + + +multiclass T_LoadAbsReg_Pat { + def : Pat <(VT (ldOp (add (shl IntRegs:$src1, u2ImmPred:$src2), + (HexagonCONST32 tglobaladdr:$src3)))), + (MI IntRegs:$src1, u2ImmPred:$src2, tglobaladdr:$src3)>; + def : Pat <(VT (ldOp (add IntRegs:$src1, + (HexagonCONST32 tglobaladdr:$src2)))), + (MI IntRegs:$src1, 0, tglobaladdr:$src2)>; + + def : Pat <(VT (ldOp (add (shl IntRegs:$src1, u2ImmPred:$src2), + (HexagonCONST32 tconstpool:$src3)))), + (MI IntRegs:$src1, u2ImmPred:$src2, tconstpool:$src3)>; + def : Pat <(VT (ldOp (add IntRegs:$src1, + (HexagonCONST32 tconstpool:$src2)))), + (MI IntRegs:$src1, 0, tconstpool:$src2)>; + + def : Pat <(VT (ldOp (add (shl IntRegs:$src1, u2ImmPred:$src2), + (HexagonCONST32 tjumptable:$src3)))), + (MI IntRegs:$src1, u2ImmPred:$src2, tjumptable:$src3)>; + def : Pat <(VT (ldOp (add IntRegs:$src1, + (HexagonCONST32 tjumptable:$src2)))), + (MI IntRegs:$src1, 0, tjumptable:$src2)>; +} + +let AddedComplexity = 60 in { +defm : T_LoadAbsReg_Pat ; +defm : T_LoadAbsReg_Pat ; +defm : T_LoadAbsReg_Pat ; + +defm : T_LoadAbsReg_Pat ; +defm : T_LoadAbsReg_Pat ; +defm : T_LoadAbsReg_Pat ; + +defm : T_LoadAbsReg_Pat ; +defm : T_LoadAbsReg_Pat ; +} //===----------------------------------------------------------------------===// // Template classes for the non-predicated load instructions with @@ -430,217 +612,234 @@ multiclass ld_idxd_shl ; defm loadrub : ld_idxd_shl<"memub", "LDriub", IntRegs, 0b001>; } -let hasNewValue = 1, accessSize = HalfWordAccess, isCodeGenOnly = 0 in { +let hasNewValue = 1, accessSize = HalfWordAccess in { defm loadrh : ld_idxd_shl<"memh", "LDrih", IntRegs, 0b010>; defm loadruh : ld_idxd_shl<"memuh", "LDriuh", IntRegs, 0b011>; } -let hasNewValue = 1, accessSize = WordAccess, isCodeGenOnly = 0 in +let hasNewValue = 1, accessSize = WordAccess in defm loadri : ld_idxd_shl<"memw", "LDriw", IntRegs, 0b100>; -let accessSize = DoubleWordAccess, isCodeGenOnly = 0 in +let accessSize = DoubleWordAccess in defm loadrd : ld_idxd_shl<"memd", "LDrid", DoubleRegs, 0b110>; // 'def pats' for load instructions with base + register offset and non-zero // immediate value. Immediate value is used to left-shift the second // register operand. +class Loadxs_pat + : Pat<(VT (Load (add (i32 IntRegs:$Rs), + (i32 (shl (i32 IntRegs:$Rt), u2ImmPred:$u2))))), + (VT (MI IntRegs:$Rs, IntRegs:$Rt, imm:$u2))>; + let AddedComplexity = 40 in { -def : Pat <(i32 (sextloadi8 (add IntRegs:$src1, - (shl IntRegs:$src2, u2ImmPred:$offset)))), - (L4_loadrb_rr IntRegs:$src1, - IntRegs:$src2, u2ImmPred:$offset)>, - Requires<[HasV4T]>; - -def : Pat <(i32 (zextloadi8 (add IntRegs:$src1, - (shl IntRegs:$src2, u2ImmPred:$offset)))), - (L4_loadrub_rr IntRegs:$src1, - IntRegs:$src2, u2ImmPred:$offset)>, - Requires<[HasV4T]>; - -def : Pat <(i32 (extloadi8 (add IntRegs:$src1, - (shl IntRegs:$src2, u2ImmPred:$offset)))), - (L4_loadrub_rr IntRegs:$src1, - IntRegs:$src2, u2ImmPred:$offset)>, - Requires<[HasV4T]>; - -def : Pat <(i32 (sextloadi16 (add IntRegs:$src1, - (shl IntRegs:$src2, u2ImmPred:$offset)))), - (L4_loadrh_rr IntRegs:$src1, - IntRegs:$src2, u2ImmPred:$offset)>, - Requires<[HasV4T]>; - -def : Pat <(i32 (zextloadi16 (add IntRegs:$src1, - (shl IntRegs:$src2, u2ImmPred:$offset)))), - (L4_loadruh_rr IntRegs:$src1, - IntRegs:$src2, u2ImmPred:$offset)>, - Requires<[HasV4T]>; - -def : Pat <(i32 (extloadi16 (add IntRegs:$src1, - (shl IntRegs:$src2, u2ImmPred:$offset)))), - (L4_loadruh_rr IntRegs:$src1, - IntRegs:$src2, u2ImmPred:$offset)>, - Requires<[HasV4T]>; - -def : Pat <(i32 (load (add IntRegs:$src1, - (shl IntRegs:$src2, u2ImmPred:$offset)))), - (L4_loadri_rr IntRegs:$src1, - IntRegs:$src2, u2ImmPred:$offset)>, - Requires<[HasV4T]>; - -def : Pat <(i64 (load (add IntRegs:$src1, - (shl IntRegs:$src2, u2ImmPred:$offset)))), - (L4_loadrd_rr IntRegs:$src1, - IntRegs:$src2, u2ImmPred:$offset)>, - Requires<[HasV4T]>; + def: Loadxs_pat; + def: Loadxs_pat; + def: Loadxs_pat; + def: Loadxs_pat; + def: Loadxs_pat; + def: Loadxs_pat; + def: Loadxs_pat; + def: Loadxs_pat; } - // 'def pats' for load instruction base + register offset and // zero immediate value. -let AddedComplexity = 10 in { -def : Pat <(i64 (load (add IntRegs:$src1, IntRegs:$src2))), - (L4_loadrd_rr IntRegs:$src1, IntRegs:$src2, 0)>, - Requires<[HasV4T]>; +class Loadxs_simple_pat + : Pat<(VT (Load (add (i32 IntRegs:$Rs), (i32 IntRegs:$Rt)))), + (VT (MI IntRegs:$Rs, IntRegs:$Rt, 0))>; + +let AddedComplexity = 20 in { + def: Loadxs_simple_pat; + def: Loadxs_simple_pat; + def: Loadxs_simple_pat; + def: Loadxs_simple_pat; + def: Loadxs_simple_pat; + def: Loadxs_simple_pat; + def: Loadxs_simple_pat; + def: Loadxs_simple_pat; +} -def : Pat <(i32 (sextloadi8 (add IntRegs:$src1, IntRegs:$src2))), - (L4_loadrb_rr IntRegs:$src1, IntRegs:$src2, 0)>, - Requires<[HasV4T]>; +// zext i1->i64 +def: Pat<(i64 (zext (i1 PredRegs:$src1))), + (Zext64 (C2_muxii PredRegs:$src1, 1, 0))>; + +// zext i32->i64 +def: Pat<(i64 (zext (i32 IntRegs:$src1))), + (Zext64 IntRegs:$src1)>; -def : Pat <(i32 (zextloadi8 (add IntRegs:$src1, IntRegs:$src2))), - (L4_loadrub_rr IntRegs:$src1, IntRegs:$src2, 0)>, - Requires<[HasV4T]>; +//===----------------------------------------------------------------------===// +// LD - +//===----------------------------------------------------------------------===// -def : Pat <(i32 (extloadi8 (add IntRegs:$src1, IntRegs:$src2))), - (L4_loadrub_rr IntRegs:$src1, IntRegs:$src2, 0)>, - Requires<[HasV4T]>; +//===----------------------------------------------------------------------===// +// ST + +//===----------------------------------------------------------------------===// +/// +//===----------------------------------------------------------------------===// +// Template class for store instructions with Absolute set addressing mode. +//===----------------------------------------------------------------------===// +let isExtended = 1, opExtendable = 1, opExtentBits = 6, + addrMode = AbsoluteSet, isNVStorable = 1 in +class T_ST_absset MajOp, MemAccessSize AccessSz, bit isHalf = 0> + : STInst<(outs IntRegs:$dst), + (ins u6Ext:$addr, RC:$src), + mnemonic#"($dst = #$addr) = $src"#!if(isHalf, ".h","")>, NewValueRel { + bits<5> dst; + bits<6> addr; + bits<5> src; + let accessSize = AccessSz; + let BaseOpcode = BaseOp#"_AbsSet"; -def : Pat <(i32 (sextloadi16 (add IntRegs:$src1, IntRegs:$src2))), - (L4_loadrh_rr IntRegs:$src1, IntRegs:$src2, 0)>, - Requires<[HasV4T]>; + let IClass = 0b1010; -def : Pat <(i32 (zextloadi16 (add IntRegs:$src1, IntRegs:$src2))), - (L4_loadruh_rr IntRegs:$src1, IntRegs:$src2, 0)>, - Requires<[HasV4T]>; + let Inst{27-24} = 0b1011; + let Inst{23-21} = MajOp; + let Inst{20-16} = dst; + let Inst{13} = 0b0; + let Inst{12-8} = src; + let Inst{7} = 0b1; + let Inst{5-0} = addr; + } -def : Pat <(i32 (extloadi16 (add IntRegs:$src1, IntRegs:$src2))), - (L4_loadruh_rr IntRegs:$src1, IntRegs:$src2, 0)>, - Requires<[HasV4T]>; +def S4_storerb_ap : T_ST_absset <"memb", "STrib", IntRegs, 0b000, ByteAccess>; +def S4_storerh_ap : T_ST_absset <"memh", "STrih", IntRegs, 0b010, + HalfWordAccess>; +def S4_storeri_ap : T_ST_absset <"memw", "STriw", IntRegs, 0b100, WordAccess>; -def : Pat <(i32 (load (add IntRegs:$src1, IntRegs:$src2))), - (L4_loadri_rr IntRegs:$src1, IntRegs:$src2, 0)>, - Requires<[HasV4T]>; +let isNVStorable = 0 in { + def S4_storerf_ap : T_ST_absset <"memh", "STrif", IntRegs, + 0b011, HalfWordAccess, 1>; + def S4_storerd_ap : T_ST_absset <"memd", "STrid", DoubleRegs, + 0b110, DoubleWordAccess>; } -// zext i1->i64 -def : Pat <(i64 (zext (i1 PredRegs:$src1))), - (i64 (A4_combineir 0, (C2_muxii (i1 PredRegs:$src1), 1, 0)))>, - Requires<[HasV4T]>; +let opExtendable = 1, isNewValue = 1, isNVStore = 1, opNewValue = 2, +isExtended = 1, opExtentBits= 6 in +class T_ST_absset_nv MajOp, + MemAccessSize AccessSz > + : NVInst <(outs IntRegs:$dst), + (ins u6Ext:$addr, IntRegs:$src), + mnemonic#"($dst = #$addr) = $src.new">, NewValueRel { + bits<5> dst; + bits<6> addr; + bits<3> src; + let accessSize = AccessSz; + let BaseOpcode = BaseOp#"_AbsSet"; -// zext i32->i64 -def : Pat <(i64 (zext (i32 IntRegs:$src1))), - (i64 (A4_combineir 0, (i32 IntRegs:$src1)))>, - Requires<[HasV4T]>; -// zext i8->i64 -def: Pat <(i64 (zextloadi8 ADDRriS11_0:$src1)), - (i64 (A4_combineir 0, (L2_loadrub_io AddrFI:$src1, 0)))>, - Requires<[HasV4T]>; - -let AddedComplexity = 20 in -def: Pat <(i64 (zextloadi8 (add (i32 IntRegs:$src1), - s11_0ExtPred:$offset))), - (i64 (A4_combineir 0, (L2_loadrub_io IntRegs:$src1, - s11_0ExtPred:$offset)))>, - Requires<[HasV4T]>; + let IClass = 0b1010; -// zext i1->i64 -def: Pat <(i64 (zextloadi1 ADDRriS11_0:$src1)), - (i64 (A4_combineir 0, (L2_loadrub_io AddrFI:$src1, 0)))>, - Requires<[HasV4T]>; - -let AddedComplexity = 20 in -def: Pat <(i64 (zextloadi1 (add (i32 IntRegs:$src1), - s11_0ExtPred:$offset))), - (i64 (A4_combineir 0, (L2_loadrub_io IntRegs:$src1, - s11_0ExtPred:$offset)))>, - Requires<[HasV4T]>; - -// zext i16->i64 -def: Pat <(i64 (zextloadi16 ADDRriS11_1:$src1)), - (i64 (A4_combineir 0, (L2_loadruh_io AddrFI:$src1, 0)))>, - Requires<[HasV4T]>; - -let AddedComplexity = 20 in -def: Pat <(i64 (zextloadi16 (add (i32 IntRegs:$src1), - s11_1ExtPred:$offset))), - (i64 (A4_combineir 0, (L2_loadruh_io IntRegs:$src1, - s11_1ExtPred:$offset)))>, - Requires<[HasV4T]>; - -// anyext i16->i64 -def: Pat <(i64 (extloadi16 ADDRriS11_2:$src1)), - (i64 (A4_combineir 0, (L2_loadrh_io AddrFI:$src1, 0)))>, - Requires<[HasV4T]>; - -let AddedComplexity = 20 in -def: Pat <(i64 (extloadi16 (add (i32 IntRegs:$src1), - s11_1ExtPred:$offset))), - (i64 (A4_combineir 0, (L2_loadrh_io IntRegs:$src1, - s11_1ExtPred:$offset)))>, - Requires<[HasV4T]>; + let Inst{27-21} = 0b1011101; + let Inst{20-16} = dst; + let Inst{13-11} = 0b000; + let Inst{12-11} = MajOp; + let Inst{10-8} = src; + let Inst{7} = 0b1; + let Inst{5-0} = addr; + } -// zext i32->i64 -def: Pat <(i64 (zextloadi32 ADDRriS11_2:$src1)), - (i64 (A4_combineir 0, (L2_loadri_io AddrFI:$src1, 0)))>, - Requires<[HasV4T]>; +let mayStore = 1, addrMode = AbsoluteSet in { + def S4_storerbnew_ap : T_ST_absset_nv <"memb", "STrib", 0b00, ByteAccess>; + def S4_storerhnew_ap : T_ST_absset_nv <"memh", "STrih", 0b01, HalfWordAccess>; + def S4_storerinew_ap : T_ST_absset_nv <"memw", "STriw", 0b10, WordAccess>; +} -let AddedComplexity = 100 in -def: Pat <(i64 (zextloadi32 (i32 (add IntRegs:$src1, s11_2ExtPred:$offset)))), - (i64 (A4_combineir 0, (L2_loadri_io IntRegs:$src1, - s11_2ExtPred:$offset)))>, - Requires<[HasV4T]>; +let isExtended = 1, opExtendable = 2, opExtentBits = 6, InputType = "imm", +addrMode = BaseLongOffset, AddedComplexity = 40 in +class T_StoreAbsReg MajOp, MemAccessSize AccessSz, bit isHalf = 0> + : STInst<(outs), + (ins IntRegs:$src1, u2Imm:$src2, u6Ext:$src3, RC:$src4), + mnemonic#"($src1<<#$src2 + #$src3) = $src4"#!if(isHalf, ".h",""), + []>, ImmRegShl, NewValueRel { -// anyext i32->i64 -def: Pat <(i64 (extloadi32 ADDRriS11_2:$src1)), - (i64 (A4_combineir 0, (L2_loadri_io AddrFI:$src1, 0)))>, - Requires<[HasV4T]>; + bits<5> src1; + bits<2> src2; + bits<6> src3; + bits<5> src4; -let AddedComplexity = 100 in -def: Pat <(i64 (extloadi32 (i32 (add IntRegs:$src1, s11_2ExtPred:$offset)))), - (i64 (A4_combineir 0, (L2_loadri_io IntRegs:$src1, - s11_2ExtPred:$offset)))>, - Requires<[HasV4T]>; + let accessSize = AccessSz; + let CextOpcode = CextOp; + let BaseOpcode = CextOp#"_shl"; + let IClass = 0b1010; + let Inst{27-24} =0b1101; + let Inst{23-21} = MajOp; + let Inst{20-16} = src1; + let Inst{13} = src2{1}; + let Inst{12-8} = src4; + let Inst{7} = 0b1; + let Inst{6} = src2{0}; + let Inst{5-0} = src3; +} +def S4_storerb_ur : T_StoreAbsReg <"memb", "STrib", IntRegs, 0b000, ByteAccess>; +def S4_storerh_ur : T_StoreAbsReg <"memh", "STrih", IntRegs, 0b010, + HalfWordAccess>; +def S4_storerf_ur : T_StoreAbsReg <"memh", "STrif", IntRegs, 0b011, + HalfWordAccess, 1>; +def S4_storeri_ur : T_StoreAbsReg <"memw", "STriw", IntRegs, 0b100, WordAccess>; +def S4_storerd_ur : T_StoreAbsReg <"memd", "STrid", DoubleRegs, 0b110, + DoubleWordAccess>; -//===----------------------------------------------------------------------===// -// LD - -//===----------------------------------------------------------------------===// +let AddedComplexity = 40 in +multiclass T_StoreAbsReg_Pats { + def : Pat<(stOp (VT RC:$src4), + (add (shl (i32 IntRegs:$src1), u2ImmPred:$src2), + u32ImmPred:$src3)), + (MI IntRegs:$src1, u2ImmPred:$src2, u32ImmPred:$src3, RC:$src4)>; -//===----------------------------------------------------------------------===// -// ST + -//===----------------------------------------------------------------------===// -/// -//===----------------------------------------------------------------------===// -// Template class for store instructions with Absolute set addressing mode. -//===----------------------------------------------------------------------===// -let isExtended = 1, opExtendable = 2, validSubTargets = HasV4SubT, -addrMode = AbsoluteSet in -class T_ST_abs_set: - STInst2<(outs IntRegs:$dst1), - (ins RC:$src1, u0AlwaysExt:$src2), - mnemonic#"($dst1=##$src2) = $src1", - []>, - Requires<[HasV4T]>; + def : Pat<(stOp (VT RC:$src4), + (add (shl IntRegs:$src1, u2ImmPred:$src2), + (HexagonCONST32 tglobaladdr:$src3))), + (MI IntRegs:$src1, u2ImmPred:$src2, tglobaladdr:$src3, RC:$src4)>; + + def : Pat<(stOp (VT RC:$src4), + (add IntRegs:$src1, (HexagonCONST32 tglobaladdr:$src3))), + (MI IntRegs:$src1, 0, tglobaladdr:$src3, RC:$src4)>; +} + +defm : T_StoreAbsReg_Pats ; +defm : T_StoreAbsReg_Pats ; +defm : T_StoreAbsReg_Pats ; +defm : T_StoreAbsReg_Pats ; + +let mayStore = 1, isNVStore = 1, isExtended = 1, addrMode = BaseLongOffset, + opExtentBits = 6, isNewValue = 1, opNewValue = 3, opExtendable = 2 in +class T_StoreAbsRegNV MajOp, + MemAccessSize AccessSz> + : NVInst <(outs ), + (ins IntRegs:$src1, u2Imm:$src2, u6Ext:$src3, IntRegs:$src4), + mnemonic#"($src1<<#$src2 + #$src3) = $src4.new">, NewValueRel { + bits<5> src1; + bits<2> src2; + bits<6> src3; + bits<3> src4; + + let CextOpcode = CextOp; + let BaseOpcode = CextOp#"_shl"; + let IClass = 0b1010; + + let Inst{27-21} = 0b1101101; + let Inst{12-11} = 0b00; + let Inst{7} = 0b1; + let Inst{20-16} = src1; + let Inst{13} = src2{1}; + let Inst{12-11} = MajOp; + let Inst{10-8} = src4; + let Inst{6} = src2{0}; + let Inst{5-0} = src3; + } -def STrid_abs_set_V4 : T_ST_abs_set <"memd", DoubleRegs>; -def STrib_abs_set_V4 : T_ST_abs_set <"memb", IntRegs>; -def STrih_abs_set_V4 : T_ST_abs_set <"memh", IntRegs>; -def STriw_abs_set_V4 : T_ST_abs_set <"memw", IntRegs>; +def S4_storerbnew_ur : T_StoreAbsRegNV <"memb", "STrib", 0b00, ByteAccess>; +def S4_storerhnew_ur : T_StoreAbsRegNV <"memh", "STrih", 0b01, HalfWordAccess>; +def S4_storerinew_ur : T_StoreAbsRegNV <"memw", "STriw", 0b10, WordAccess>; //===----------------------------------------------------------------------===// // Template classes for the non-predicated store instructions with @@ -804,8 +1003,7 @@ multiclass ST_Idxd_shl_nv , ST_Idxd_shl_nv<"memb", "STrib", IntRegs, 0b00>; @@ -825,83 +1023,18 @@ let addrMode = BaseRegOffset, InputType = "reg", hasSideEffects = 0, defm storerf: ST_Idxd_shl<"memh", "STrif", IntRegs, 0b011, 1>; } -let Predicates = [HasV4T], AddedComplexity = 10 in { -def : Pat<(truncstorei8 (i32 IntRegs:$src4), - (add IntRegs:$src1, (shl IntRegs:$src2, - u2ImmPred:$src3))), - (S4_storerb_rr IntRegs:$src1, IntRegs:$src2, - u2ImmPred:$src3, IntRegs:$src4)>; - -def : Pat<(truncstorei16 (i32 IntRegs:$src4), - (add IntRegs:$src1, (shl IntRegs:$src2, - u2ImmPred:$src3))), - (S4_storerh_rr IntRegs:$src1, IntRegs:$src2, - u2ImmPred:$src3, IntRegs:$src4)>; - -def : Pat<(store (i32 IntRegs:$src4), - (add IntRegs:$src1, (shl IntRegs:$src2, u2ImmPred:$src3))), - (S4_storeri_rr IntRegs:$src1, IntRegs:$src2, - u2ImmPred:$src3, IntRegs:$src4)>; - -def : Pat<(store (i64 DoubleRegs:$src4), - (add IntRegs:$src1, (shl IntRegs:$src2, u2ImmPred:$src3))), - (S4_storerd_rr IntRegs:$src1, IntRegs:$src2, - u2ImmPred:$src3, DoubleRegs:$src4)>; -} - -let isExtended = 1, opExtendable = 2 in -class T_ST_LongOff : - STInst<(outs), - (ins IntRegs:$src1, u2Imm:$src2, u0AlwaysExt:$src3, RC:$src4), - mnemonic#"($src1<<#$src2+##$src3) = $src4", - [(stOp (VT RC:$src4), - (add (shl (i32 IntRegs:$src1), u2ImmPred:$src2), - u0AlwaysExtPred:$src3))]>, - Requires<[HasV4T]>; - -let isExtended = 1, opExtendable = 2, mayStore = 1, isNVStore = 1 in -class T_ST_LongOff_nv : - NVInst_V4<(outs), - (ins IntRegs:$src1, u2Imm:$src2, u0AlwaysExt:$src3, IntRegs:$src4), - mnemonic#"($src1<<#$src2+##$src3) = $src4.new", - []>, - Requires<[HasV4T]>; - -multiclass ST_LongOff { - let BaseOpcode = BaseOp#"_shl" in { - let isNVStorable = 1 in - def NAME#_V4 : T_ST_LongOff; - - def NAME#_nv_V4 : T_ST_LongOff_nv; - } -} - -let AddedComplexity = 10, validSubTargets = HasV4SubT in { - def STrid_shl_V4 : T_ST_LongOff<"memd", store, DoubleRegs, i64>; - defm STrib_shl : ST_LongOff <"memb", "STrib", truncstorei8>, NewValueRel; - defm STrih_shl : ST_LongOff <"memh", "Strih", truncstorei16>, NewValueRel; - defm STriw_shl : ST_LongOff <"memw", "STriw", store>, NewValueRel; -} - -let AddedComplexity = 40 in -multiclass T_ST_LOff_Pats { - def : Pat<(stOp (VT RC:$src4), - (add (shl IntRegs:$src1, u2ImmPred:$src2), - (NumUsesBelowThresCONST32 tglobaladdr:$src3))), - (I IntRegs:$src1, u2ImmPred:$src2, tglobaladdr:$src3, RC:$src4)>; +class Storexs_pat + : Pat<(Store Value:$Ru, (add (i32 IntRegs:$Rs), + (i32 (shl (i32 IntRegs:$Rt), u2ImmPred:$u2)))), + (MI IntRegs:$Rs, IntRegs:$Rt, imm:$u2, Value:$Ru)>; - def : Pat<(stOp (VT RC:$src4), - (add IntRegs:$src1, - (NumUsesBelowThresCONST32 tglobaladdr:$src3))), - (I IntRegs:$src1, 0, tglobaladdr:$src3, RC:$src4)>; +let AddedComplexity = 40 in { + def: Storexs_pat; + def: Storexs_pat; + def: Storexs_pat; + def: Storexs_pat; } -defm : T_ST_LOff_Pats; -defm : T_ST_LOff_Pats; -defm : T_ST_LOff_Pats; -defm : T_ST_LOff_Pats; - // memd(Rx++#s4:3)=Rtt // memd(Rx++#s4:3:circ(Mu))=Rtt // memd(Rx++I:circ(Mu))=Rtt @@ -1004,8 +1137,8 @@ multiclass ST_Imm ; @@ -1016,22 +1149,49 @@ let hasSideEffects = 0, validSubTargets = HasV4SubT, addrMode = BaseImmOffset, defm S4_storeiri : ST_Imm<"memw", "STriw", u6_2Imm, 0b10>; } -let Predicates = [HasV4T], AddedComplexity = 10 in { -def: Pat<(truncstorei8 s8ExtPred:$src3, (add IntRegs:$src1, u6_0ImmPred:$src2)), - (S4_storeirb_io IntRegs:$src1, u6_0ImmPred:$src2, s8ExtPred:$src3)>; +def IMM_BYTE : SDNodeXFormgetSExtValue(); + return CurDAG->getTargetConstant(imm, SDLoc(N), MVT::i32); +}]>; + +def IMM_HALF : SDNodeXFormgetSExtValue(); + return CurDAG->getTargetConstant(imm, SDLoc(N), MVT::i32); +}]>; + +def IMM_WORD : SDNodeXFormgetSExtValue(); + return CurDAG->getTargetConstant(imm, SDLoc(N), MVT::i32); +}]>; -def: Pat<(truncstorei16 s8ExtPred:$src3, (add IntRegs:$src1, - u6_1ImmPred:$src2)), - (S4_storeirh_io IntRegs:$src1, u6_1ImmPred:$src2, s8ExtPred:$src3)>; +def ToImmByte : OutPatFrag<(ops node:$R), (IMM_BYTE $R)>; +def ToImmHalf : OutPatFrag<(ops node:$R), (IMM_HALF $R)>; +def ToImmWord : OutPatFrag<(ops node:$R), (IMM_WORD $R)>; -def: Pat<(store s8ExtPred:$src3, (add IntRegs:$src1, u6_2ImmPred:$src2)), - (S4_storeiri_io IntRegs:$src1, u6_2ImmPred:$src2, s8ExtPred:$src3)>; +let AddedComplexity = 40 in { + // Not using frameindex patterns for these stores, because the offset + // is not extendable. This could cause problems during removing the frame + // indices, since the offset with respect to R29/R30 may not fit in the + // u6 field. + def: Storexm_add_pat; + def: Storexm_add_pat; + def: Storexm_add_pat; } -let AddedComplexity = 6 in -def : Pat <(truncstorei8 s8ExtPred:$src2, (i32 IntRegs:$src1)), - (S4_storeirb_io IntRegs:$src1, 0, s8ExtPred:$src2)>, - Requires<[HasV4T]>; +def: Storexm_simple_pat; +def: Storexm_simple_pat; +def: Storexm_simple_pat; // memb(Rx++#s4:0:circ(Mu))=Rt // memb(Rx++I:circ(Mu))=Rt @@ -1039,16 +1199,10 @@ def : Pat <(truncstorei8 s8ExtPred:$src2, (i32 IntRegs:$src1)), // memb(Rx++Mu:brev)=Rt // memb(gp+#u16:0)=Rt - // Store halfword. // TODO: needs to be implemented // memh(Re=#U6)=Rt.H // memh(Rs+#s11:1)=Rt.H -let AddedComplexity = 6 in -def : Pat <(truncstorei16 s8ExtPred:$src2, (i32 IntRegs:$src1)), - (S4_storeirh_io IntRegs:$src1, 0, s8ExtPred:$src2)>, - Requires<[HasV4T]>; - // memh(Rs+Ru<<#u2)=Rt.H // TODO: needs to be implemented. @@ -1065,7 +1219,6 @@ def : Pat <(truncstorei16 s8ExtPred:$src2, (i32 IntRegs:$src1)), // if ([!]Pv[.new]) memh(#u6)=Rt.H // if ([!]Pv[.new]) memh(#u6)=Rt - // if ([!]Pv[.new]) memh(Rs+#u6:1)=Rt.H // TODO: needs to be implemented. @@ -1075,20 +1228,6 @@ def : Pat <(truncstorei16 s8ExtPred:$src2, (i32 IntRegs:$src1)), // Store word. // memw(Re=#U6)=Rt // TODO: Needs to be implemented. - -// Store predicate: -let hasSideEffects = 0 in -def STriw_pred_V4 : STInst2<(outs), - (ins MEMri:$addr, PredRegs:$src1), - "Error; should not emit", - []>, - Requires<[HasV4T]>; - -let AddedComplexity = 6 in -def : Pat <(store s8ExtPred:$src2, (i32 IntRegs:$src1)), - (S4_storeiri_io IntRegs:$src1, 0, s8ExtPred:$src2)>, - Requires<[HasV4T]>; - // memw(Rx++#s4:2)=Rt // memw(Rx++#s4:2:circ(Mu))=Rt // memw(Rx++I:circ(Mu))=Rt @@ -1203,7 +1342,7 @@ multiclass ST_Idxd_nv, AddrModeRel; @@ -1217,12 +1356,46 @@ let addrMode = BaseImmOffset, InputType = "imm", isCodeGenOnly = 0 in { u6_2Ext, 0b10>, AddrModeRel; } +//===----------------------------------------------------------------------===// +// Post increment loads with register offset. +//===----------------------------------------------------------------------===// + +let hasNewValue = 1 in +def L2_loadbsw2_pr : T_load_pr <"membh", IntRegs, 0b0001, HalfWordAccess>; + +def L2_loadbsw4_pr : T_load_pr <"membh", DoubleRegs, 0b0111, WordAccess>; + +let hasSideEffects = 0, addrMode = PostInc in +class T_loadalign_pr MajOp, MemAccessSize AccessSz> + : LDInstPI <(outs DoubleRegs:$dst, IntRegs:$_dst_), + (ins DoubleRegs:$src1, IntRegs:$src2, ModRegs:$src3), + "$dst = "#mnemonic#"($src2++$src3)", [], + "$src1 = $dst, $src2 = $_dst_"> { + bits<5> dst; + bits<5> src2; + bits<1> src3; + + let accessSize = AccessSz; + let IClass = 0b1001; + + let Inst{27-25} = 0b110; + let Inst{24-21} = MajOp; + let Inst{20-16} = src2; + let Inst{13} = src3; + let Inst{12} = 0b0; + let Inst{7} = 0b0; + let Inst{4-0} = dst; + } + +def L2_loadalignb_pr : T_loadalign_pr <"memb_fifo", 0b0100, ByteAccess>; +def L2_loadalignh_pr : T_loadalign_pr <"memh_fifo", 0b0010, HalfWordAccess>; + //===----------------------------------------------------------------------===// // Template class for non-predicated post increment .new stores // mem[bhwd](Rx++#s4:[0123])=Nt.new //===----------------------------------------------------------------------===// -let isPredicable = 1, hasSideEffects = 0, validSubTargets = HasV4SubT, - addrMode = PostInc, isNVStore = 1, isNewValue = 1, opNewValue = 3 in +let isPredicable = 1, hasSideEffects = 0, addrMode = PostInc, isNVStore = 1, + isNewValue = 1, opNewValue = 3 in class T_StorePI_nv MajOp > : NVInstPI_V4 <(outs IntRegs:$_dst_), (ins IntRegs:$src1, ImmOp:$offset, IntRegs:$src2), @@ -1254,8 +1427,8 @@ class T_StorePI_nv MajOp > // Template class for predicated post increment .new stores // if([!]Pv[.new]) mem[bhwd](Rx++#s4:[0123])=Nt.new //===----------------------------------------------------------------------===// -let isPredicated = 1, hasSideEffects = 0, validSubTargets = HasV4SubT, - addrMode = PostInc, isNVStore = 1, isNewValue = 1, opNewValue = 4 in +let isPredicated = 1, hasSideEffects = 0, addrMode = PostInc, isNVStore = 1, + isNewValue = 1, opNewValue = 4 in class T_StorePI_nv_pred MajOp, bit isPredNot, bit isPredNew > : NVInstPI_V4 <(outs IntRegs:$_dst_), @@ -1310,13 +1483,13 @@ multiclass ST_PostInc_nv; -let accessSize = HalfWordAccess, isCodeGenOnly = 0 in +let accessSize = HalfWordAccess in defm storerhnew: ST_PostInc_nv <"memh", "STrih", s4_1Imm, 0b01>; -let accessSize = WordAccess, isCodeGenOnly = 0 in +let accessSize = WordAccess in defm storerinew: ST_PostInc_nv <"memw", "STriw", s4_2Imm, 0b10>; //===----------------------------------------------------------------------===// @@ -1343,15 +1516,12 @@ class T_StorePI_RegNV MajOp, MemAccessSize AccessSz> let Inst{7} = 0b0; } -let isCodeGenOnly = 0 in { def S2_storerbnew_pr : T_StorePI_RegNV<"memb", 0b00, ByteAccess>; def S2_storerhnew_pr : T_StorePI_RegNV<"memh", 0b01, HalfWordAccess>; def S2_storerinew_pr : T_StorePI_RegNV<"memw", 0b10, WordAccess>; -} // memb(Rx++#s4:0:circ(Mu))=Nt.new // memb(Rx++I:circ(Mu))=Nt.new -// memb(Rx++Mu)=Nt.new // memb(Rx++Mu:brev)=Nt.new // memh(Rx++#s4:1:circ(Mu))=Nt.new // memh(Rx++I:circ(Mu))=Nt.new @@ -1401,7 +1571,7 @@ class NVJrr_template majOp, bit NvOpNum, let RegOp = !if(!eq(NvOpNum, 0), src2, src1); let IClass = 0b0010; - let Inst{26} = 0b0; + let Inst{27-26} = 0b00; let Inst{25-23} = majOp; let Inst{22} = isNegCond; let Inst{18-16} = Ns; @@ -1415,9 +1585,9 @@ class NVJrr_template majOp, bit NvOpNum, multiclass NVJrr_cond majOp, bit NvOpNum, bit isNegCond> { // Branch not taken: - def _nt_V4: NVJrr_template; + def _nt: NVJrr_template; // Branch taken: - def _t_V4: NVJrr_template; + def _t : NVJrr_template; } // NvOpNum = 0 -> First Operand is a new-value Register @@ -1426,8 +1596,8 @@ multiclass NVJrr_cond majOp, bit NvOpNum, multiclass NVJrr_base majOp, bit NvOpNum> { let BaseOpcode = BaseOp#_NVJ in { - defm _t_Jumpnv : NVJrr_cond; // True cond - defm _f_Jumpnv : NVJrr_cond; // False cond + defm _t_jumpnv : NVJrr_cond; // True cond + defm _f_jumpnv : NVJrr_cond; // False cond } } @@ -1438,13 +1608,12 @@ multiclass NVJrr_base majOp, // if ([!]cmp.gtu(Rt,Ns.new)) jump:[n]t #r9:2 let isPredicated = 1, isBranch = 1, isNewValue = 1, isTerminator = 1, - Defs = [PC], hasSideEffects = 0, validSubTargets = HasV4SubT, - isCodeGenOnly = 0 in { - defm CMPEQrr : NVJrr_base<"cmp.eq", "CMPEQ", 0b000, 0>, PredRel; - defm CMPGTrr : NVJrr_base<"cmp.gt", "CMPGT", 0b001, 0>, PredRel; - defm CMPGTUrr : NVJrr_base<"cmp.gtu", "CMPGTU", 0b010, 0>, PredRel; - defm CMPLTrr : NVJrr_base<"cmp.gt", "CMPLT", 0b011, 1>, PredRel; - defm CMPLTUrr : NVJrr_base<"cmp.gtu", "CMPLTU", 0b100, 1>, PredRel; + Defs = [PC], hasSideEffects = 0 in { + defm J4_cmpeq : NVJrr_base<"cmp.eq", "CMPEQ", 0b000, 0>, PredRel; + defm J4_cmpgt : NVJrr_base<"cmp.gt", "CMPGT", 0b001, 0>, PredRel; + defm J4_cmpgtu : NVJrr_base<"cmp.gtu", "CMPGTU", 0b010, 0>, PredRel; + defm J4_cmplt : NVJrr_base<"cmp.gt", "CMPLT", 0b011, 1>, PredRel; + defm J4_cmpltu : NVJrr_base<"cmp.gtu", "CMPLTU", 0b100, 1>, PredRel; } //===----------------------------------------------------------------------===// @@ -1482,15 +1651,15 @@ class NVJri_template majOp, bit isNegCond, multiclass NVJri_cond majOp, bit isNegCond> { // Branch not taken: - def _nt_V4: NVJri_template; + def _nt: NVJri_template; // Branch taken: - def _t_V4: NVJri_template; + def _t : NVJri_template; } multiclass NVJri_base majOp> { let BaseOpcode = BaseOp#_NVJri in { - defm _t_Jumpnv : NVJri_cond; // True Cond - defm _f_Jumpnv : NVJri_cond; // False cond + defm _t_jumpnv : NVJri_cond; // True Cond + defm _f_jumpnv : NVJri_cond; // False cond } } @@ -1499,11 +1668,10 @@ multiclass NVJri_base majOp> { // if ([!]cmp.gtu(Ns.new,#U5)) jump:[n]t #r9:2 let isPredicated = 1, isBranch = 1, isNewValue = 1, isTerminator = 1, - Defs = [PC], hasSideEffects = 0, validSubTargets = HasV4SubT, - isCodeGenOnly = 0 in { - defm CMPEQri : NVJri_base<"cmp.eq", "CMPEQ", 0b000>, PredRel; - defm CMPGTri : NVJri_base<"cmp.gt", "CMPGT", 0b001>, PredRel; - defm CMPGTUri : NVJri_base<"cmp.gtu", "CMPGTU", 0b010>, PredRel; + Defs = [PC], hasSideEffects = 0 in { + defm J4_cmpeqi : NVJri_base<"cmp.eq", "CMPEQ", 0b000>, PredRel; + defm J4_cmpgti : NVJri_base<"cmp.gt", "CMPGT", 0b001>, PredRel; + defm J4_cmpgtui : NVJri_base<"cmp.gtu", "CMPGTU", 0b010>, PredRel; } //===----------------------------------------------------------------------===// @@ -1540,16 +1708,16 @@ class NVJ_ConstImm_template majOp, string ImmVal, multiclass NVJ_ConstImm_cond majOp, string ImmVal, bit isNegCond> { // Branch not taken: - def _nt_V4: NVJ_ConstImm_template; + def _nt: NVJ_ConstImm_template; // Branch taken: - def _t_V4: NVJ_ConstImm_template; + def _t : NVJ_ConstImm_template; } multiclass NVJ_ConstImm_base majOp, string ImmVal> { let BaseOpcode = BaseOp#_NVJ_ConstImm in { - defm _t_Jumpnv : NVJ_ConstImm_cond; // True - defm _f_Jumpnv : NVJ_ConstImm_cond; // False + defm _t_jumpnv : NVJ_ConstImm_cond; // True + defm _f_jumpnv : NVJ_ConstImm_cond; // False } } @@ -1558,14 +1726,14 @@ multiclass NVJ_ConstImm_base majOp, // if ([!]cmp.gt(Ns.new,#-1)) jump:[n]t #r9:2 let isPredicated = 1, isBranch = 1, isNewValue = 1, isTerminator=1, - Defs = [PC], hasSideEffects = 0, isCodeGenOnly = 0 in { - defm TSTBIT0 : NVJ_ConstImm_base<"tstbit", "TSTBIT", 0b011, "0">, PredRel; - defm CMPEQn1 : NVJ_ConstImm_base<"cmp.eq", "CMPEQ", 0b100, "-1">, PredRel; - defm CMPGTn1 : NVJ_ConstImm_base<"cmp.gt", "CMPGT", 0b101, "-1">, PredRel; + Defs = [PC], hasSideEffects = 0 in { + defm J4_tstbit0 : NVJ_ConstImm_base<"tstbit", "TSTBIT", 0b011, "0">, PredRel; + defm J4_cmpeqn1 : NVJ_ConstImm_base<"cmp.eq", "CMPEQ", 0b100, "-1">, PredRel; + defm J4_cmpgtn1 : NVJ_ConstImm_base<"cmp.gt", "CMPGT", 0b101, "-1">, PredRel; } // J4_hintjumpr: Hint indirect conditional jump. -let isBranch = 1, isIndirectBranch = 1, hasSideEffects = 0, isCodeGenOnly = 0 in +let isBranch = 1, isIndirectBranch = 1, hasSideEffects = 0 in def J4_hintjumpr: JRInst < (outs), (ins IntRegs:$Rs), @@ -1586,8 +1754,7 @@ def J4_hintjumpr: JRInst < // PC-relative add let hasNewValue = 1, isExtendable = 1, opExtendable = 1, - isExtentSigned = 0, opExtentBits = 6, hasSideEffects = 0, - Uses = [PC], validSubTargets = HasV4SubT in + isExtentSigned = 0, opExtentBits = 6, hasSideEffects = 0, Uses = [PC] in def C4_addipc : CRInst <(outs IntRegs:$Rd), (ins u6Ext:$u6), "$Rd = add(pc, #$u6)", [], "", CR_tc_2_SLOT3 > { bits<5> Rd; @@ -1625,7 +1792,6 @@ class T_LOGICAL_3OP OpBits, bit IsNeg> let Inst{1-0} = Pd; } -let isCodeGenOnly = 0 in { def C4_and_and : T_LOGICAL_3OP<"and", "and", 0b00, 0>; def C4_and_or : T_LOGICAL_3OP<"and", "or", 0b01, 0>; def C4_or_and : T_LOGICAL_3OP<"or", "and", 0b10, 0>; @@ -1634,7 +1800,69 @@ def C4_and_andn : T_LOGICAL_3OP<"and", "and", 0b00, 1>; def C4_and_orn : T_LOGICAL_3OP<"and", "or", 0b01, 1>; def C4_or_andn : T_LOGICAL_3OP<"or", "and", 0b10, 1>; def C4_or_orn : T_LOGICAL_3OP<"or", "or", 0b11, 1>; -} + +// op(Ps, op(Pt, Pu)) +class LogLog_pat + : Pat<(i1 (Op1 I1:$Ps, (Op2 I1:$Pt, I1:$Pu))), + (MI I1:$Ps, I1:$Pt, I1:$Pu)>; + +// op(Ps, op(Pt, ~Pu)) +class LogLogNot_pat + : Pat<(i1 (Op1 I1:$Ps, (Op2 I1:$Pt, (not I1:$Pu)))), + (MI I1:$Ps, I1:$Pt, I1:$Pu)>; + +def: LogLog_pat; +def: LogLog_pat; +def: LogLog_pat; +def: LogLog_pat; + +def: LogLogNot_pat; +def: LogLogNot_pat; +def: LogLogNot_pat; +def: LogLogNot_pat; + +//===----------------------------------------------------------------------===// +// PIC: Support for PIC compilations. The patterns and SD nodes defined +// below are needed to support code generation for PIC +//===----------------------------------------------------------------------===// + +def SDT_HexagonPICAdd + : SDTypeProfile<1, 1, [SDTCisVT<0, i32>, SDTCisVT<1, i32>]>; +def SDT_HexagonGOTAdd + : SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisVT<1, i32>]>; + +def SDT_HexagonGOTAddInternal : SDTypeProfile<1, 1, [SDTCisVT<0, i32>]>; +def SDT_HexagonGOTAddInternalJT : SDTypeProfile<1, 1, [SDTCisVT<0, i32>]>; +def SDT_HexagonGOTAddInternalBA : SDTypeProfile<1, 1, [SDTCisVT<0, i32>]>; + +def Hexagonpic_add : SDNode<"HexagonISD::PIC_ADD", SDT_HexagonPICAdd>; +def Hexagonat_got : SDNode<"HexagonISD::AT_GOT", SDT_HexagonGOTAdd>; +def Hexagongat_pcrel : SDNode<"HexagonISD::AT_PCREL", + SDT_HexagonGOTAddInternal>; +def Hexagongat_pcrel_jt : SDNode<"HexagonISD::AT_PCREL", + SDT_HexagonGOTAddInternalJT>; +def Hexagongat_pcrel_ba : SDNode<"HexagonISD::AT_PCREL", + SDT_HexagonGOTAddInternalBA>; + +// PIC: Map from a block address computation to a PC-relative add +def: Pat<(Hexagongat_pcrel_ba tblockaddress:$src1), + (C4_addipc u32ImmPred:$src1)>; + +// PIC: Map from the computation to generate a GOT pointer to a PC-relative add +def: Pat<(Hexagonpic_add texternalsym:$src1), + (C4_addipc u32ImmPred:$src1)>; + +// PIC: Map from a jump table address computation to a PC-relative add +def: Pat<(Hexagongat_pcrel_jt tjumptable:$src1), + (C4_addipc u32ImmPred:$src1)>; + +// PIC: Map from a GOT-relative symbol reference to a load +def: Pat<(Hexagonat_got (i32 IntRegs:$src1), tglobaladdr:$src2), + (L2_loadri_io IntRegs:$src1, s30_2ImmPred:$src2)>; + +// PIC: Map from a static symbol reference to a PC-relative add +def: Pat<(Hexagongat_pcrel tglobaladdr:$src1), + (C4_addipc u32ImmPred:$src1)>; //===----------------------------------------------------------------------===// // CR - @@ -1645,12 +1873,15 @@ def C4_or_orn : T_LOGICAL_3OP<"or", "or", 0b11, 1>; //===----------------------------------------------------------------------===// // Logical with-not instructions. -let validSubTargets = HasV4SubT, isCodeGenOnly = 0 in { - def A4_andnp : T_ALU64_logical<"and", 0b001, 1, 0, 1>; - def A4_ornp : T_ALU64_logical<"or", 0b011, 1, 0, 1>; -} +def A4_andnp : T_ALU64_logical<"and", 0b001, 1, 0, 1>; +def A4_ornp : T_ALU64_logical<"or", 0b011, 1, 0, 1>; + +def: Pat<(i64 (and (i64 DoubleRegs:$Rs), (i64 (not (i64 DoubleRegs:$Rt))))), + (A4_andnp DoubleRegs:$Rs, DoubleRegs:$Rt)>; +def: Pat<(i64 (or (i64 DoubleRegs:$Rs), (i64 (not (i64 DoubleRegs:$Rt))))), + (A4_ornp DoubleRegs:$Rs, DoubleRegs:$Rt)>; -let hasNewValue = 1, hasSideEffects = 0, isCodeGenOnly = 0 in +let hasNewValue = 1, hasSideEffects = 0 in def S4_parity: ALU64Inst<(outs IntRegs:$Rd), (ins IntRegs:$Rs, IntRegs:$Rt), "$Rd = parity($Rs, $Rt)", [], "", ALU64_tc_2_SLOT23> { bits<5> Rd; @@ -1663,15 +1894,16 @@ def S4_parity: ALU64Inst<(outs IntRegs:$Rd), (ins IntRegs:$Rs, IntRegs:$Rt), let Inst{12-8} = Rt; let Inst{4-0} = Rd; } + // Add and accumulate. // Rd=add(Rs,add(Ru,#s6)) let isExtentSigned = 1, hasNewValue = 1, isExtendable = 1, opExtentBits = 6, - opExtendable = 3, isCodeGenOnly = 0 in + opExtendable = 3 in def S4_addaddi : ALU64Inst <(outs IntRegs:$Rd), (ins IntRegs:$Rs, IntRegs:$Ru, s6Ext:$s6), "$Rd = add($Rs, add($Ru, #$s6))" , [(set (i32 IntRegs:$Rd), (add (i32 IntRegs:$Rs), - (add (i32 IntRegs:$Ru), s6_16ExtPred:$s6)))], + (add (i32 IntRegs:$Ru), s16_16ImmPred:$s6)))], "", ALU64_tc_2_SLOT23> { bits<5> Rd; bits<5> Rs; @@ -1690,7 +1922,7 @@ def S4_addaddi : ALU64Inst <(outs IntRegs:$Rd), } let isExtentSigned = 1, hasSideEffects = 0, hasNewValue = 1, isExtendable = 1, - opExtentBits = 6, opExtendable = 2, isCodeGenOnly = 0 in + opExtentBits = 6, opExtendable = 2 in def S4_subaddi: ALU64Inst <(outs IntRegs:$Rd), (ins IntRegs:$Rs, s6Ext:$s6, IntRegs:$Ru), "$Rd = add($Rs, sub(#$s6, $Ru))", @@ -1710,31 +1942,64 @@ def S4_subaddi: ALU64Inst <(outs IntRegs:$Rd), let Inst{7-5} = s6{2-0}; let Inst{4-0} = Ru; } - + +// Rd=add(Rs,sub(#s6,Ru)) +def: Pat<(add (i32 IntRegs:$src1), (sub s32ImmPred:$src2, + (i32 IntRegs:$src3))), + (S4_subaddi IntRegs:$src1, s32ImmPred:$src2, IntRegs:$src3)>; + +// Rd=sub(add(Rs,#s6),Ru) +def: Pat<(sub (add (i32 IntRegs:$src1), s32ImmPred:$src2), + (i32 IntRegs:$src3)), + (S4_subaddi IntRegs:$src1, s32ImmPred:$src2, IntRegs:$src3)>; + +// Rd=add(sub(Rs,Ru),#s6) +def: Pat<(add (sub (i32 IntRegs:$src1), (i32 IntRegs:$src3)), + (s32ImmPred:$src2)), + (S4_subaddi IntRegs:$src1, s32ImmPred:$src2, IntRegs:$src3)>; + + +// Add or subtract doublewords with carry. +//TODO: +// Rdd=add(Rss,Rtt,Px):carry +//TODO: +// Rdd=sub(Rss,Rtt,Px):carry + // Extract bitfield // Rdd=extract(Rss,#u6,#U6) // Rdd=extract(Rss,Rtt) // Rd=extract(Rs,Rtt) // Rd=extract(Rs,#u5,#U5) -let isCodeGenOnly = 0 in { def S4_extractp_rp : T_S3op_64 < "extract", 0b11, 0b100, 0>; def S4_extractp : T_S2op_extract <"extract", 0b1010, DoubleRegs, u6Imm>; -} -let hasNewValue = 1, isCodeGenOnly = 0 in { +let hasNewValue = 1 in { def S4_extract_rp : T_S3op_extract<"extract", 0b01>; def S4_extract : T_S2op_extract <"extract", 0b1101, IntRegs, u5Imm>; } -let Itinerary = M_tc_3x_SLOT23, Defs = [USR_OVF], isCodeGenOnly = 0 in { +// Complex add/sub halfwords/words +let Defs = [USR_OVF] in { + def S4_vxaddsubh : T_S3op_64 < "vxaddsubh", 0b01, 0b100, 0, 1>; + def S4_vxaddsubw : T_S3op_64 < "vxaddsubw", 0b01, 0b000, 0, 1>; + def S4_vxsubaddh : T_S3op_64 < "vxsubaddh", 0b01, 0b110, 0, 1>; + def S4_vxsubaddw : T_S3op_64 < "vxsubaddw", 0b01, 0b010, 0, 1>; +} + +let Defs = [USR_OVF] in { + def S4_vxaddsubhr : T_S3op_64 < "vxaddsubh", 0b11, 0b000, 0, 1, 1, 1>; + def S4_vxsubaddhr : T_S3op_64 < "vxsubaddh", 0b11, 0b010, 0, 1, 1, 1>; +} + +let Itinerary = M_tc_3x_SLOT23, Defs = [USR_OVF] in { def M4_mac_up_s1_sat: T_MType_acc_rr<"+= mpy", 0b011, 0b000, 0, [], 0, 1, 1>; def M4_nac_up_s1_sat: T_MType_acc_rr<"-= mpy", 0b011, 0b001, 0, [], 0, 1, 1>; } // Logical xor with xor accumulation. // Rxx^=xor(Rss,Rtt) -let hasSideEffects = 0, isCodeGenOnly = 0 in +let hasSideEffects = 0 in def M4_xor_xacc : SInst <(outs DoubleRegs:$Rxx), (ins DoubleRegs:$dst2, DoubleRegs:$Rss, DoubleRegs:$Rtt), @@ -1749,40 +2014,106 @@ def M4_xor_xacc let IClass = 0b1100; - let Inst{27-23} = 0b10101; + let Inst{27-22} = 0b101010; let Inst{20-16} = Rss; let Inst{12-8} = Rtt; + let Inst{7-5} = 0b000; let Inst{4-0} = Rxx; } - -// Split bitfield -let isCodeGenOnly = 0 in -def A4_bitspliti : T_S2op_2_di <"bitsplit", 0b110, 0b100>; -// Arithmetic/Convergent round -let isCodeGenOnly = 0 in -def A4_cround_ri : T_S2op_2_ii <"cround", 0b111, 0b000>; +// Rotate and reduce bytes +// Rdd=vrcrotate(Rss,Rt,#u2) +let hasSideEffects = 0 in +def S4_vrcrotate + : SInst <(outs DoubleRegs:$Rdd), + (ins DoubleRegs:$Rss, IntRegs:$Rt, u2Imm:$u2), + "$Rdd = vrcrotate($Rss, $Rt, #$u2)", + [], "", S_3op_tc_3x_SLOT23> { + bits<5> Rdd; + bits<5> Rss; + bits<5> Rt; + bits<2> u2; -let isCodeGenOnly = 0 in -def A4_round_ri : T_S2op_2_ii <"round", 0b111, 0b100>; + let IClass = 0b1100; -let Defs = [USR_OVF], isCodeGenOnly = 0 in -def A4_round_ri_sat : T_S2op_2_ii <"round", 0b111, 0b110, 1>; + let Inst{27-22} = 0b001111; + let Inst{20-16} = Rss; + let Inst{13} = u2{1}; + let Inst{12-8} = Rt; + let Inst{7-6} = 0b11; + let Inst{5} = u2{0}; + let Inst{4-0} = Rdd; + } -// Logical-logical words. -// Compound or-and -- Rx=or(Ru,and(Rx,#s10)) -let isExtentSigned = 1, hasNewValue = 1, isExtendable = 1, opExtentBits = 10, - opExtendable = 3, isCodeGenOnly = 0 in -def S4_or_andix: - ALU64Inst<(outs IntRegs:$Rx), - (ins IntRegs:$Ru, IntRegs:$_src_, s10Ext:$s10), - "$Rx = or($Ru, and($_src_, #$s10))" , - [(set (i32 IntRegs:$Rx), - (or (i32 IntRegs:$Ru), (and (i32 IntRegs:$_src_), s10ExtPred:$s10)))] , - "$_src_ = $Rx", ALU64_tc_2_SLOT23> { - bits<5> Rx; - bits<5> Ru; - bits<10> s10; +// Rotate and reduce bytes with accumulation +// Rxx+=vrcrotate(Rss,Rt,#u2) +let hasSideEffects = 0 in +def S4_vrcrotate_acc + : SInst <(outs DoubleRegs:$Rxx), + (ins DoubleRegs:$dst2, DoubleRegs:$Rss, IntRegs:$Rt, u2Imm:$u2), + "$Rxx += vrcrotate($Rss, $Rt, #$u2)", [], + "$dst2 = $Rxx", S_3op_tc_3x_SLOT23> { + bits<5> Rxx; + bits<5> Rss; + bits<5> Rt; + bits<2> u2; + + let IClass = 0b1100; + + let Inst{27-21} = 0b1011101; + let Inst{20-16} = Rss; + let Inst{13} = u2{1}; + let Inst{12-8} = Rt; + let Inst{5} = u2{0}; + let Inst{4-0} = Rxx; + } + +// Vector reduce conditional negate halfwords +let hasSideEffects = 0 in +def S2_vrcnegh + : SInst <(outs DoubleRegs:$Rxx), + (ins DoubleRegs:$dst2, DoubleRegs:$Rss, IntRegs:$Rt), + "$Rxx += vrcnegh($Rss, $Rt)", [], + "$dst2 = $Rxx", S_3op_tc_3x_SLOT23> { + bits<5> Rxx; + bits<5> Rss; + bits<5> Rt; + + let IClass = 0b1100; + + let Inst{27-21} = 0b1011001; + let Inst{20-16} = Rss; + let Inst{13} = 0b1; + let Inst{12-8} = Rt; + let Inst{7-5} = 0b111; + let Inst{4-0} = Rxx; + } + +// Split bitfield +def A4_bitspliti : T_S2op_2_di <"bitsplit", 0b110, 0b100>; + +// Arithmetic/Convergent round +def A4_cround_ri : T_S2op_2_ii <"cround", 0b111, 0b000>; + +def A4_round_ri : T_S2op_2_ii <"round", 0b111, 0b100>; + +let Defs = [USR_OVF] in +def A4_round_ri_sat : T_S2op_2_ii <"round", 0b111, 0b110, 1>; + +// Logical-logical words. +// Compound or-and -- Rx=or(Ru,and(Rx,#s10)) +let isExtentSigned = 1, hasNewValue = 1, isExtendable = 1, opExtentBits = 10, + opExtendable = 3 in +def S4_or_andix: + ALU64Inst<(outs IntRegs:$Rx), + (ins IntRegs:$Ru, IntRegs:$_src_, s10Ext:$s10), + "$Rx = or($Ru, and($_src_, #$s10))" , + [(set (i32 IntRegs:$Rx), + (or (i32 IntRegs:$Ru), (and (i32 IntRegs:$_src_), s32ImmPred:$s10)))] , + "$_src_ = $Rx", ALU64_tc_2_SLOT23> { + bits<5> Rx; + bits<5> Ru; + bits<10> s10; let IClass = 0b1101; @@ -1795,7 +2126,7 @@ def S4_or_andix: // Miscellaneous ALU64 instructions. // -let hasNewValue = 1, hasSideEffects = 0, isCodeGenOnly = 0 in +let hasNewValue = 1, hasSideEffects = 0 in def A4_modwrapu: ALU64Inst<(outs IntRegs:$Rd), (ins IntRegs:$Rs, IntRegs:$Rt), "$Rd = modwrap($Rs, $Rt)", [], "", ALU64_tc_2_SLOT23> { bits<5> Rd; @@ -1810,7 +2141,7 @@ def A4_modwrapu: ALU64Inst<(outs IntRegs:$Rd), (ins IntRegs:$Rs, IntRegs:$Rt), let Inst{4-0} = Rd; } -let hasSideEffects = 0, isCodeGenOnly = 0 in +let hasSideEffects = 0 in def A4_bitsplit: ALU64Inst<(outs DoubleRegs:$Rd), (ins IntRegs:$Rs, IntRegs:$Rt), "$Rd = bitsplit($Rs, $Rt)", [], "", ALU64_tc_1_SLOT23> { @@ -1826,7 +2157,54 @@ def A4_bitsplit: ALU64Inst<(outs DoubleRegs:$Rd), let Inst{4-0} = Rd; } -let isCodeGenOnly = 0 in { +let hasSideEffects = 0 in +def dep_S2_packhl: ALU64Inst<(outs DoubleRegs:$Rd), + (ins IntRegs:$Rs, IntRegs:$Rt), + "$Rd = packhl($Rs, $Rt):deprecated", [], "", ALU64_tc_1_SLOT23> { + bits<5> Rd; + bits<5> Rs; + bits<5> Rt; + + let IClass = 0b1101; + let Inst{27-24} = 0b0100; + let Inst{21} = 0b0; + let Inst{20-16} = Rs; + let Inst{12-8} = Rt; + let Inst{4-0} = Rd; +} + +let hasNewValue = 1, hasSideEffects = 0 in +def dep_A2_addsat: ALU64Inst<(outs IntRegs:$Rd), + (ins IntRegs:$Rs, IntRegs:$Rt), + "$Rd = add($Rs, $Rt):sat:deprecated", [], "", ALU64_tc_2_SLOT23> { + bits<5> Rd; + bits<5> Rs; + bits<5> Rt; + + let IClass = 0b1101; + let Inst{27-21} = 0b0101100; + let Inst{20-16} = Rs; + let Inst{12-8} = Rt; + let Inst{7} = 0b0; + let Inst{4-0} = Rd; +} + +let hasNewValue = 1, hasSideEffects = 0 in +def dep_A2_subsat: ALU64Inst<(outs IntRegs:$Rd), + (ins IntRegs:$Rs, IntRegs:$Rt), + "$Rd = sub($Rs, $Rt):sat:deprecated", [], "", ALU64_tc_2_SLOT23> { + bits<5> Rd; + bits<5> Rs; + bits<5> Rt; + + let IClass = 0b1101; + let Inst{27-21} = 0b0101100; + let Inst{20-16} = Rt; + let Inst{12-8} = Rs; + let Inst{7} = 0b1; + let Inst{4-0} = Rd; +} + // Rx[&|]=xor(Rs,Rt) def M4_or_xor : T_MType_acc_rr < "|= xor", 0b110, 0b001, 0>; def M4_and_xor : T_MType_acc_rr < "&= xor", 0b010, 0b010, 0>; @@ -1849,7 +2227,24 @@ def M4_and_and : T_MType_acc_rr < "&= and", 0b010, 0b000, 0>; def M4_xor_andn : T_MType_acc_rr < "^= and", 0b001, 0b010, 0, [], 1>; def M4_or_andn : T_MType_acc_rr < "|= and", 0b001, 0b000, 0, [], 1>; def M4_and_andn : T_MType_acc_rr < "&= and", 0b001, 0b001, 0, [], 1>; -} + +def: T_MType_acc_pat2 ; +def: T_MType_acc_pat2 ; +def: T_MType_acc_pat2 ; +def: T_MType_acc_pat2 ; +def: T_MType_acc_pat2 ; +def: T_MType_acc_pat2 ; +def: T_MType_acc_pat2 ; +def: T_MType_acc_pat2 ; + +class T_MType_acc_pat3 + : Pat <(i32 (secOp IntRegs:$src1, (firstOp IntRegs:$src2, + (not IntRegs:$src3)))), + (i32 (MI IntRegs:$src1, IntRegs:$src2, IntRegs:$src3))>; + +def: T_MType_acc_pat3 ; +def: T_MType_acc_pat3 ; +def: T_MType_acc_pat3 ; // Compound or-or and or-and let isExtentSigned = 1, InputType = "imm", hasNewValue = 1, isExtendable = 1, @@ -1859,7 +2254,7 @@ class T_CompOR MajOp, SDNode OpNode> (ins IntRegs:$src1, IntRegs:$Rs, s10Ext:$s10), "$Rx |= "#mnemonic#"($Rs, #$s10)", [(set (i32 IntRegs:$Rx), (or (i32 IntRegs:$src1), - (OpNode (i32 IntRegs:$Rs), s10ExtPred:$s10)))], + (OpNode (i32 IntRegs:$Rs), s32ImmPred:$s10)))], "$src1 = $Rx", ALU64_tc_2_SLOT23>, ImmRegRel { bits<5> Rx; bits<5> Rs; @@ -1875,10 +2270,10 @@ class T_CompOR MajOp, SDNode OpNode> let Inst{4-0} = Rx; } -let CextOpcode = "ORr_ANDr", isCodeGenOnly = 0 in +let CextOpcode = "ORr_ANDr" in def S4_or_andi : T_CompOR <"and", 0b00, and>; -let CextOpcode = "ORr_ORr", isCodeGenOnly = 0 in +let CextOpcode = "ORr_ORr" in def S4_or_ori : T_CompOR <"or", 0b10, or>; // Modulo wrap @@ -1923,22 +2318,33 @@ def S4_or_ori : T_CompOR <"or", 0b10, or>; //===----------------------------------------------------------------------===// // Bit reverse -let isCodeGenOnly = 0 in def S2_brevp : T_S2op_3 <"brev", 0b11, 0b110>; // Bit count -let isCodeGenOnly = 0 in { def S2_ct0p : T_COUNT_LEADING_64<"ct0", 0b111, 0b010>; def S2_ct1p : T_COUNT_LEADING_64<"ct1", 0b111, 0b100>; def S4_clbpnorm : T_COUNT_LEADING_64<"normamt", 0b011, 0b000>; -} -def: Pat<(i32 (trunc (cttz (i64 DoubleRegs:$Rss)))), - (S2_ct0p (i64 DoubleRegs:$Rss))>; -def: Pat<(i32 (trunc (cttz (not (i64 DoubleRegs:$Rss))))), - (S2_ct1p (i64 DoubleRegs:$Rss))>; +// Count trailing zeros: 64-bit. +def: Pat<(i32 (trunc (cttz I64:$Rss))), (S2_ct0p I64:$Rss)>; +def: Pat<(i32 (trunc (cttz_zero_undef I64:$Rss))), (S2_ct0p I64:$Rss)>; + +// Count trailing ones: 64-bit. +def: Pat<(i32 (trunc (cttz (not I64:$Rss)))), (S2_ct1p I64:$Rss)>; +def: Pat<(i32 (trunc (cttz_zero_undef (not I64:$Rss)))), (S2_ct1p I64:$Rss)>; -let hasSideEffects = 0, hasNewValue = 1, isCodeGenOnly = 0 in +// Define leading/trailing patterns that require zero-extensions to 64 bits. +def: Pat<(i64 (ctlz I64:$Rss)), (Zext64 (S2_cl0p I64:$Rss))>; +def: Pat<(i64 (ctlz_zero_undef I64:$Rss)), (Zext64 (S2_cl0p I64:$Rss))>; +def: Pat<(i64 (cttz I64:$Rss)), (Zext64 (S2_ct0p I64:$Rss))>; +def: Pat<(i64 (cttz_zero_undef I64:$Rss)), (Zext64 (S2_ct0p I64:$Rss))>; +def: Pat<(i64 (ctlz (not I64:$Rss))), (Zext64 (S2_cl1p I64:$Rss))>; +def: Pat<(i64 (ctlz_zero_undef (not I64:$Rss))), (Zext64 (S2_cl1p I64:$Rss))>; +def: Pat<(i64 (cttz (not I64:$Rss))), (Zext64 (S2_ct1p I64:$Rss))>; +def: Pat<(i64 (cttz_zero_undef (not I64:$Rss))), (Zext64 (S2_ct1p I64:$Rss))>; + + +let hasSideEffects = 0, hasNewValue = 1 in def S4_clbaddi : SInst<(outs IntRegs:$Rd), (ins IntRegs:$Rs, s6Imm:$s6), "$Rd = add(clb($Rs), #$s6)", [], "", S_2op_tc_2_SLOT23> { bits<5> Rs; @@ -1953,7 +2359,7 @@ def S4_clbaddi : SInst<(outs IntRegs:$Rd), (ins IntRegs:$Rs, s6Imm:$s6), let Inst{4-0} = Rd; } -let hasSideEffects = 0, hasNewValue = 1, isCodeGenOnly = 0 in +let hasSideEffects = 0, hasNewValue = 1 in def S4_clbpaddi : SInst<(outs IntRegs:$Rd), (ins DoubleRegs:$Rs, s6Imm:$s6), "$Rd = add(clb($Rs), #$s6)", [], "", S_2op_tc_2_SLOT23> { bits<5> Rs; @@ -1970,10 +2376,8 @@ def S4_clbpaddi : SInst<(outs IntRegs:$Rd), (ins DoubleRegs:$Rs, s6Imm:$s6), // Bit test/set/clear -let isCodeGenOnly = 0 in { def S4_ntstbit_i : T_TEST_BIT_IMM<"!tstbit", 0b001>; def S4_ntstbit_r : T_TEST_BIT_REG<"!tstbit", 1>; -} let AddedComplexity = 20 in { // Complexity greater than cmp reg-imm. def: Pat<(i1 (seteq (and (shl 1, u5ImmPred:$u5), (i32 IntRegs:$Rs)), 0)), @@ -1993,11 +2397,9 @@ let AddedComplexity = 100 in def: Pat<(i1 (seteq (and (i32 IntRegs:$Rs), (i32 Set5ImmPred:$u5)), (i32 0))), (S4_ntstbit_i (i32 IntRegs:$Rs), (BITPOS32 Set5ImmPred:$u5))>; -let isCodeGenOnly = 0 in { def C4_nbitsset : T_TEST_BITS_REG<"!bitsset", 0b01, 1>; def C4_nbitsclr : T_TEST_BITS_REG<"!bitsclr", 0b10, 1>; def C4_nbitsclri : T_TEST_BITS_IMM<"!bitsclr", 0b10, 1>; -} // Do not increase complexity of these patterns. In the DAG, "cmp i8" may be // represented as a compare against "value & 0xFF", which is an exact match @@ -2022,14 +2424,13 @@ def: Pat<(i1 (setne (and I32:$Rs, I32:$Rt), I32:$Rt)), // Rd=add(#u6,mpyi(Rs,#U6)) -- Multiply by immed and add immed. -let hasNewValue = 1, isExtendable = 1, opExtentBits = 6, opExtendable = 1, - isCodeGenOnly = 0 in +let hasNewValue = 1, isExtendable = 1, opExtentBits = 6, opExtendable = 1 in def M4_mpyri_addi : MInst<(outs IntRegs:$Rd), (ins u6Ext:$u6, IntRegs:$Rs, u6Imm:$U6), "$Rd = add(#$u6, mpyi($Rs, #$U6))" , [(set (i32 IntRegs:$Rd), (add (mul (i32 IntRegs:$Rs), u6ImmPred:$U6), - u6ExtPred:$u6))] ,"",ALU64_tc_3x_SLOT23> { + u32ImmPred:$u6))] ,"",ALU64_tc_3x_SLOT23> { bits<5> Rd; bits<6> u6; bits<5> Rs; @@ -2049,12 +2450,12 @@ def M4_mpyri_addi : MInst<(outs IntRegs:$Rd), // Rd=add(#u6,mpyi(Rs,Rt)) let CextOpcode = "ADD_MPY", InputType = "imm", hasNewValue = 1, - isExtendable = 1, opExtentBits = 6, opExtendable = 1, isCodeGenOnly = 0 in + isExtendable = 1, opExtentBits = 6, opExtendable = 1 in def M4_mpyrr_addi : MInst <(outs IntRegs:$Rd), (ins u6Ext:$u6, IntRegs:$Rs, IntRegs:$Rt), "$Rd = add(#$u6, mpyi($Rs, $Rt))" , [(set (i32 IntRegs:$Rd), - (add (mul (i32 IntRegs:$Rs), (i32 IntRegs:$Rt)), u6ExtPred:$u6))], + (add (mul (i32 IntRegs:$Rs), (i32 IntRegs:$Rt)), u32ImmPred:$u6))], "", ALU64_tc_3x_SLOT23>, ImmRegRel { bits<5> Rd; bits<6> u6; @@ -2099,18 +2500,16 @@ class T_AddMpy let Inst{4-0} = src1; } -let isCodeGenOnly = 0 in def M4_mpyri_addr_u2 : T_AddMpy<0b0, u6_2ImmPred, (ins IntRegs:$src1, u6_2Imm:$src2, IntRegs:$src3)>; let isExtendable = 1, opExtentBits = 6, opExtendable = 3, - CextOpcode = "ADD_MPY", InputType = "imm", isCodeGenOnly = 0 in -def M4_mpyri_addr : T_AddMpy<0b1, u6ExtPred, + CextOpcode = "ADD_MPY", InputType = "imm" in +def M4_mpyri_addr : T_AddMpy<0b1, u32ImmPred, (ins IntRegs:$src1, IntRegs:$src3, u6Ext:$src2)>, ImmRegRel; // Rx=add(Ru,mpyi(Rx,Rs)) -let validSubTargets = HasV4SubT, CextOpcode = "ADD_MPY", InputType = "reg", - hasNewValue = 1, isCodeGenOnly = 0 in +let CextOpcode = "ADD_MPY", InputType = "reg", hasNewValue = 1 in def M4_mpyrr_addr: MInst_acc <(outs IntRegs:$Rx), (ins IntRegs:$Ru, IntRegs:$_src_, IntRegs:$Rs), "$Rx = add($Ru, mpyi($_src_, $Rs))", @@ -2129,51 +2528,101 @@ def M4_mpyrr_addr: MInst_acc <(outs IntRegs:$Rx), let Inst{20-16} = Rs; } -// Rd=add(##,mpyi(Rs,#U6)) -def : Pat <(add (mul (i32 IntRegs:$src2), u6ImmPred:$src3), - (HexagonCONST32 tglobaladdr:$src1)), - (i32 (M4_mpyri_addi tglobaladdr:$src1, IntRegs:$src2, - u6ImmPred:$src3))>; -// Rd=add(##,mpyi(Rs,Rt)) -def : Pat <(add (mul (i32 IntRegs:$src2), (i32 IntRegs:$src3)), - (HexagonCONST32 tglobaladdr:$src1)), - (i32 (M4_mpyrr_addi tglobaladdr:$src1, IntRegs:$src2, - IntRegs:$src3))>; +// Vector reduce multiply word by signed half (32x16) +//Rdd=vrmpyweh(Rss,Rtt)[:<<1] +def M4_vrmpyeh_s0 : T_M2_vmpy<"vrmpyweh", 0b010, 0b100, 0, 0, 0>; +def M4_vrmpyeh_s1 : T_M2_vmpy<"vrmpyweh", 0b110, 0b100, 1, 0, 0>; -// Polynomial multiply words -// Rdd=pmpyw(Rs,Rt) -// Rxx^=pmpyw(Rs,Rt) +//Rdd=vrmpywoh(Rss,Rtt)[:<<1] +def M4_vrmpyoh_s0 : T_M2_vmpy<"vrmpywoh", 0b001, 0b010, 0, 0, 0>; +def M4_vrmpyoh_s1 : T_M2_vmpy<"vrmpywoh", 0b101, 0b010, 1, 0, 0>; -// Vector reduce multiply word by signed half (32x16) -// Rdd=vrmpyweh(Rss,Rtt)[:<<1] -// Rdd=vrmpywoh(Rss,Rtt)[:<<1] -// Rxx+=vrmpyweh(Rss,Rtt)[:<<1] -// Rxx+=vrmpywoh(Rss,Rtt)[:<<1] - -// Multiply and use upper result -// Rd=mpy(Rs,Rt.H):<<1:sat -// Rd=mpy(Rs,Rt.L):<<1:sat -// Rd=mpy(Rs,Rt):<<1 -// Rd=mpy(Rs,Rt):<<1:sat -// Rd=mpysu(Rs,Rt) -// Rx+=mpy(Rs,Rt):<<1:sat -// Rx-=mpy(Rs,Rt):<<1:sat - -// Vector multiply bytes -// Rdd=vmpybsu(Rs,Rt) -// Rdd=vmpybu(Rs,Rt) -// Rxx+=vmpybsu(Rs,Rt) -// Rxx+=vmpybu(Rs,Rt) +//Rdd+=vrmpyweh(Rss,Rtt)[:<<1] +def M4_vrmpyeh_acc_s0: T_M2_vmpy_acc<"vrmpyweh", 0b001, 0b110, 0, 0>; +def M4_vrmpyeh_acc_s1: T_M2_vmpy_acc<"vrmpyweh", 0b101, 0b110, 1, 0>; + +//Rdd=vrmpywoh(Rss,Rtt)[:<<1] +def M4_vrmpyoh_acc_s0: T_M2_vmpy_acc<"vrmpywoh", 0b011, 0b110, 0, 0>; +def M4_vrmpyoh_acc_s1: T_M2_vmpy_acc<"vrmpywoh", 0b111, 0b110, 1, 0>; + +// Vector multiply halfwords, signed by unsigned +// Rdd=vmpyhsu(Rs,Rt)[:<<]:sat +def M2_vmpy2su_s0 : T_XTYPE_mpy64 < "vmpyhsu", 0b000, 0b111, 1, 0, 0>; +def M2_vmpy2su_s1 : T_XTYPE_mpy64 < "vmpyhsu", 0b100, 0b111, 1, 1, 0>; + +// Rxx+=vmpyhsu(Rs,Rt)[:<<1]:sat +def M2_vmac2su_s0 : T_XTYPE_mpy64_acc < "vmpyhsu", "+", 0b011, 0b101, 1, 0, 0>; +def M2_vmac2su_s1 : T_XTYPE_mpy64_acc < "vmpyhsu", "+", 0b111, 0b101, 1, 1, 0>; // Vector polynomial multiply halfwords // Rdd=vpmpyh(Rs,Rt) +def M4_vpmpyh : T_XTYPE_mpy64 < "vpmpyh", 0b110, 0b111, 0, 0, 0>; + // Rxx^=vpmpyh(Rs,Rt) +def M4_vpmpyh_acc : T_XTYPE_mpy64_acc < "vpmpyh", "^", 0b101, 0b111, 0, 0, 0>; + +// Polynomial multiply words +// Rdd=pmpyw(Rs,Rt) +def M4_pmpyw : T_XTYPE_mpy64 < "pmpyw", 0b010, 0b111, 0, 0, 0>; + +// Rxx^=pmpyw(Rs,Rt) +def M4_pmpyw_acc : T_XTYPE_mpy64_acc < "pmpyw", "^", 0b001, 0b111, 0, 0, 0>; //===----------------------------------------------------------------------===// // XTYPE/MPY - //===----------------------------------------------------------------------===// +//===----------------------------------------------------------------------===// +// ALU64/Vector compare +//===----------------------------------------------------------------------===// +//===----------------------------------------------------------------------===// +// Template class for vector compare +//===----------------------------------------------------------------------===// + +let hasSideEffects = 0 in +class T_vcmpImm cmpOp, bits<2> minOp, Operand ImmOprnd> + : ALU64_rr <(outs PredRegs:$Pd), + (ins DoubleRegs:$Rss, ImmOprnd:$Imm), + "$Pd = "#Str#"($Rss, #$Imm)", + [], "", ALU64_tc_2early_SLOT23> { + bits<2> Pd; + bits<5> Rss; + bits<32> Imm; + bits<8> ImmBits; + let ImmBits{6-0} = Imm{6-0}; + let ImmBits{7} = !if (!eq(cmpOp,0b10), 0b0, Imm{7}); // 0 for vcmp[bhw].gtu + + let IClass = 0b1101; + + let Inst{27-24} = 0b1100; + let Inst{22-21} = cmpOp; + let Inst{20-16} = Rss; + let Inst{12-5} = ImmBits; + let Inst{4-3} = minOp; + let Inst{1-0} = Pd; + } + +// Vector compare bytes +def A4_vcmpbgt : T_vcmp <"vcmpb.gt", 0b1010>; +def: T_vcmp_pat; + +let AsmString = "$Pd = any8(vcmpb.eq($Rss, $Rtt))" in +def A4_vcmpbeq_any : T_vcmp <"any8(vcmpb.gt", 0b1000>; + +def A4_vcmpbeqi : T_vcmpImm <"vcmpb.eq", 0b00, 0b00, u8Imm>; +def A4_vcmpbgti : T_vcmpImm <"vcmpb.gt", 0b01, 0b00, s8Imm>; +def A4_vcmpbgtui : T_vcmpImm <"vcmpb.gtu", 0b10, 0b00, u7Imm>; + +// Vector compare halfwords +def A4_vcmpheqi : T_vcmpImm <"vcmph.eq", 0b00, 0b01, s8Imm>; +def A4_vcmphgti : T_vcmpImm <"vcmph.gt", 0b01, 0b01, s8Imm>; +def A4_vcmphgtui : T_vcmpImm <"vcmph.gtu", 0b10, 0b01, u7Imm>; + +// Vector compare words +def A4_vcmpweqi : T_vcmpImm <"vcmpw.eq", 0b00, 0b10, s8Imm>; +def A4_vcmpwgti : T_vcmpImm <"vcmpw.gt", 0b01, 0b10, s8Imm>; +def A4_vcmpwgtui : T_vcmpImm <"vcmpw.gtu", 0b10, 0b10, u7Imm>; //===----------------------------------------------------------------------===// // XTYPE/SHIFT + @@ -2184,13 +2633,13 @@ def : Pat <(add (mul (i32 IntRegs:$src2), (i32 IntRegs:$src3)), // Rx=and(#u8,asl(Rx,#U5)) Rx=and(#u8,lsr(Rx,#U5)) // Rx=or(#u8,asl(Rx,#U5)) Rx=or(#u8,lsr(Rx,#U5)) let isExtendable = 1, opExtendable = 1, isExtentSigned = 0, opExtentBits = 8, - hasNewValue = 1, opNewValue = 0, validSubTargets = HasV4SubT in + hasNewValue = 1, opNewValue = 0 in class T_S4_ShiftOperate MajOp, InstrItinClass Itin> : MInst_acc<(outs IntRegs:$Rd), (ins u8Ext:$u8, IntRegs:$Rx, u5Imm:$U5), "$Rd = "#MnOp#"(#$u8, "#MnSh#"($Rx, #$U5))", [(set (i32 IntRegs:$Rd), - (Op (Sh I32:$Rx, u5ImmPred:$U5), u8ExtPred:$u8))], + (Op (Sh I32:$Rx, u5ImmPred:$U5), u32ImmPred:$u8))], "$Rd = $Rx", Itin> { bits<5> Rd; @@ -2216,29 +2665,48 @@ multiclass T_ShiftOperate MajOp, def _lsr_ri : T_S4_ShiftOperate; } -let AddedComplexity = 200, isCodeGenOnly = 0 in { +let AddedComplexity = 200 in { defm S4_addi : T_ShiftOperate<"add", add, 0b10, ALU64_tc_2_SLOT23>; defm S4_andi : T_ShiftOperate<"and", and, 0b00, ALU64_tc_2_SLOT23>; } -let AddedComplexity = 30, isCodeGenOnly = 0 in +let AddedComplexity = 30 in defm S4_ori : T_ShiftOperate<"or", or, 0b01, ALU64_tc_1_SLOT23>; -let isCodeGenOnly = 0 in defm S4_subi : T_ShiftOperate<"sub", sub, 0b11, ALU64_tc_1_SLOT23>; +let AddedComplexity = 200 in { + def: Pat<(add addrga:$addr, (shl I32:$src2, u5ImmPred:$src3)), + (S4_addi_asl_ri addrga:$addr, IntRegs:$src2, u5ImmPred:$src3)>; + def: Pat<(add addrga:$addr, (srl I32:$src2, u5ImmPred:$src3)), + (S4_addi_lsr_ri addrga:$addr, IntRegs:$src2, u5ImmPred:$src3)>; + def: Pat<(sub addrga:$addr, (shl I32:$src2, u5ImmPred:$src3)), + (S4_subi_asl_ri addrga:$addr, IntRegs:$src2, u5ImmPred:$src3)>; + def: Pat<(sub addrga:$addr, (srl I32:$src2, u5ImmPred:$src3)), + (S4_subi_lsr_ri addrga:$addr, IntRegs:$src2, u5ImmPred:$src3)>; +} + +// Vector conditional negate +// Rdd=vcnegh(Rss,Rt) +let Defs = [USR_OVF], Itinerary = S_3op_tc_2_SLOT23 in +def S2_vcnegh : T_S3op_shiftVect < "vcnegh", 0b11, 0b01>; // Rd=[cround|round](Rs,Rt) -let hasNewValue = 1, Itinerary = S_3op_tc_2_SLOT23, isCodeGenOnly = 0 in { +let hasNewValue = 1, Itinerary = S_3op_tc_2_SLOT23 in { def A4_cround_rr : T_S3op_3 < "cround", IntRegs, 0b11, 0b00>; def A4_round_rr : T_S3op_3 < "round", IntRegs, 0b11, 0b10>; } // Rd=round(Rs,Rt):sat -let hasNewValue = 1, Defs = [USR_OVF], Itinerary = S_3op_tc_2_SLOT23, - isCodeGenOnly = 0 in +let hasNewValue = 1, Defs = [USR_OVF], Itinerary = S_3op_tc_2_SLOT23 in def A4_round_rr_sat : T_S3op_3 < "round", IntRegs, 0b11, 0b11, 1>; +// Rd=[cmpyiwh|cmpyrwh](Rss,Rt):<<1:rnd:sat +let Defs = [USR_OVF], Itinerary = S_3op_tc_3x_SLOT23 in { + def M4_cmpyi_wh : T_S3op_8<"cmpyiwh", 0b100, 1, 1, 1>; + def M4_cmpyr_wh : T_S3op_8<"cmpyrwh", 0b110, 1, 1, 1>; +} + // Rdd=[add|sub](Rss,Rtt,Px):carry let isPredicateLate = 1, hasSideEffects = 0 in class T_S3op_carry MajOp> @@ -2261,13 +2729,51 @@ class T_S3op_carry MajOp> let Inst{4-0} = Rdd; } -let isCodeGenOnly = 0 in { def A4_addp_c : T_S3op_carry < "add", 0b110 >; def A4_subp_c : T_S3op_carry < "sub", 0b111 >; -} + +let Itinerary = S_3op_tc_3_SLOT23, hasSideEffects = 0 in +class T_S3op_6 MinOp, bit isUnsigned> + : SInst <(outs DoubleRegs:$Rxx), + (ins DoubleRegs:$dst2, DoubleRegs:$Rss, IntRegs:$Ru), + "$Rxx = "#mnemonic#"($Rss, $Ru)" , + [] , "$dst2 = $Rxx"> { + bits<5> Rxx; + bits<5> Rss; + bits<5> Ru; + + let IClass = 0b1100; + + let Inst{27-21} = 0b1011001; + let Inst{20-16} = Rss; + let Inst{13} = isUnsigned; + let Inst{12-8} = Rxx; + let Inst{7-5} = MinOp; + let Inst{4-0} = Ru; + } + +// Vector reduce maximum halfwords +// Rxx=vrmax[u]h(Rss,Ru) +def A4_vrmaxh : T_S3op_6 < "vrmaxh", 0b001, 0>; +def A4_vrmaxuh : T_S3op_6 < "vrmaxuh", 0b001, 1>; + +// Vector reduce maximum words +// Rxx=vrmax[u]w(Rss,Ru) +def A4_vrmaxw : T_S3op_6 < "vrmaxw", 0b010, 0>; +def A4_vrmaxuw : T_S3op_6 < "vrmaxuw", 0b010, 1>; + +// Vector reduce minimum halfwords +// Rxx=vrmin[u]h(Rss,Ru) +def A4_vrminh : T_S3op_6 < "vrminh", 0b101, 0>; +def A4_vrminuh : T_S3op_6 < "vrminuh", 0b101, 1>; + +// Vector reduce minimum words +// Rxx=vrmin[u]w(Rss,Ru) +def A4_vrminw : T_S3op_6 < "vrminw", 0b110, 0>; +def A4_vrminuw : T_S3op_6 < "vrminuw", 0b110, 1>; // Shift an immediate left by register amount. -let hasNewValue = 1, hasSideEffects = 0, isCodeGenOnly = 0 in +let hasNewValue = 1, hasSideEffects = 0 in def S4_lsli: SInst <(outs IntRegs:$Rd), (ins s6Imm:$s6, IntRegs:$Rt), "$Rd = lsl(#$s6, $Rt)" , [(set (i32 IntRegs:$Rd), (shl s6ImmPred:$s6, @@ -2299,7 +2805,7 @@ def MEMOPIMM : SDNodeXFormgetSExtValue(); - return XformM5ToU5Imm(imm); + return XformM5ToU5Imm(imm, SDLoc(N)); }]>; def MEMOPIMM_HALF : SDNodeXFormgetSExtValue(); - return XformM5ToU5Imm(imm); + return XformM5ToU5Imm(imm, SDLoc(N)); }]>; def MEMOPIMM_BYTE : SDNodeXFormgetSExtValue(); - return XformM5ToU5Imm(imm); + return XformM5ToU5Imm(imm, SDLoc(N)); }]>; def SETMEMIMM : SDNodeXFormgetSExtValue(); - return XformMskToBitPosU5Imm(imm); + return XformMskToBitPosU5Imm(imm, SDLoc(N)); }]>; def CLRMEMIMM : SDNodeXFormgetSExtValue()); - return XformMskToBitPosU5Imm(imm); + return XformMskToBitPosU5Imm(imm, SDLoc(N)); }]>; def SETMEMIMM_SHORT : SDNodeXFormgetSExtValue(); - return XformMskToBitPosU4Imm(imm); + return XformMskToBitPosU4Imm(imm, SDLoc(N)); }]>; def CLRMEMIMM_SHORT : SDNodeXFormgetSExtValue()); - return XformMskToBitPosU4Imm(imm); + return XformMskToBitPosU4Imm(imm, SDLoc(N)); }]>; def SETMEMIMM_BYTE : SDNodeXFormgetSExtValue(); - return XformMskToBitPosU3Imm(imm); + return XformMskToBitPosU3Imm(imm, SDLoc(N)); }]>; def CLRMEMIMM_BYTE : SDNodeXFormgetSExtValue()); - return XformMskToBitPosU3Imm(imm); + return XformMskToBitPosU3Imm(imm, SDLoc(N)); }]>; //===----------------------------------------------------------------------===// @@ -2450,15 +2956,14 @@ multiclass MemOp_base opcBits, Operand ImmOp> { } // Define MemOp instructions. -let isExtendable = 1, opExtendable = 1, isExtentSigned = 0, - validSubTargets =HasV4SubT in { - let opExtentBits = 6, accessSize = ByteAccess, isCodeGenOnly = 0 in +let isExtendable = 1, opExtendable = 1, isExtentSigned = 0 in { + let opExtentBits = 6, accessSize = ByteAccess in defm memopb_io : MemOp_base <"memb", 0b00, u6_0Ext>; - let opExtentBits = 7, accessSize = HalfWordAccess, isCodeGenOnly = 0 in + let opExtentBits = 7, accessSize = HalfWordAccess in defm memoph_io : MemOp_base <"memh", 0b01, u6_1Ext>; - let opExtentBits = 8, accessSize = WordAccess, isCodeGenOnly = 0 in + let opExtentBits = 8, accessSize = WordAccess in defm memopw_io : MemOp_base <"memw", 0b10, u6_2Ext>; } @@ -2469,43 +2974,43 @@ let isExtendable = 1, opExtendable = 1, isExtentSigned = 0, // mem[bh](Rs+#u6) += #U5 //===----------------------------------------------------------------------===// -multiclass MemOpi_u5Pats { let AddedComplexity = 180 in - def : Pat < (stOp (OpNode (ldOp IntRegs:$addr), u5ImmPred:$addend), - IntRegs:$addr), - (MI IntRegs:$addr, #0, u5ImmPred:$addend )>; + def: Pat<(stOp (OpNode (ldOp IntRegs:$addr), u5ImmPred:$addend), + IntRegs:$addr), + (MI IntRegs:$addr, 0, u5ImmPred:$addend)>; let AddedComplexity = 190 in - def : Pat <(stOp (OpNode (ldOp (add IntRegs:$base, ExtPred:$offset)), - u5ImmPred:$addend), - (add IntRegs:$base, ExtPred:$offset)), - (MI IntRegs:$base, ExtPred:$offset, u5ImmPred:$addend)>; + def: Pat<(stOp (OpNode (ldOp (add IntRegs:$base, ImmPred:$offset)), + u5ImmPred:$addend), + (add IntRegs:$base, ImmPred:$offset)), + (MI IntRegs:$base, ImmPred:$offset, u5ImmPred:$addend)>; } -multiclass MemOpi_u5ALUOp { - defm : MemOpi_u5Pats; - defm : MemOpi_u5Pats; + defm: MemOpi_u5Pats; + defm: MemOpi_u5Pats; } multiclass MemOpi_u5ExtType { // Half Word - defm : MemOpi_u5ALUOp ; + defm: MemOpi_u5ALUOp ; // Byte - defm : MemOpi_u5ALUOp ; + defm: MemOpi_u5ALUOp ; } -let Predicates = [HasV4T, UseMEMOP] in { - defm : MemOpi_u5ExtType; // zero extend - defm : MemOpi_u5ExtType; // sign extend - defm : MemOpi_u5ExtType; // any extend +let Predicates = [UseMEMOP] in { + defm: MemOpi_u5ExtType; // zero extend + defm: MemOpi_u5ExtType; // sign extend + defm: MemOpi_u5ExtType; // any extend // Word - defm : MemOpi_u5ALUOp ; + defm: MemOpi_u5ALUOp ; } //===----------------------------------------------------------------------===// @@ -2515,38 +3020,37 @@ let Predicates = [HasV4T, UseMEMOP] in { // mem[bh](Rs+#u6) += #m5 //===----------------------------------------------------------------------===// -multiclass MemOpi_m5Pats { +multiclass MemOpi_m5Pats { let AddedComplexity = 190 in - def : Pat <(stOp (add (ldOp IntRegs:$addr), immPred:$subend), - IntRegs:$addr), - (MI IntRegs:$addr, #0, (xformFunc immPred:$subend) )>; + def: Pat<(stOp (add (ldOp IntRegs:$addr), immPred:$subend), IntRegs:$addr), + (MI IntRegs:$addr, 0, (xformFunc immPred:$subend))>; let AddedComplexity = 195 in - def : Pat<(stOp (add (ldOp (add IntRegs:$base, extPred:$offset)), - immPred:$subend), - (add IntRegs:$base, extPred:$offset)), - (MI IntRegs:$base, extPred:$offset, (xformFunc immPred:$subend))>; + def: Pat<(stOp (add (ldOp (add IntRegs:$base, ImmPred:$offset)), + immPred:$subend), + (add IntRegs:$base, ImmPred:$offset)), + (MI IntRegs:$base, ImmPred:$offset, (xformFunc immPred:$subend))>; } multiclass MemOpi_m5ExtType { // Half Word - defm : MemOpi_m5Pats ; + defm: MemOpi_m5Pats ; // Byte - defm : MemOpi_m5Pats ; + defm: MemOpi_m5Pats ; } -let Predicates = [HasV4T, UseMEMOP] in { - defm : MemOpi_m5ExtType; // zero extend - defm : MemOpi_m5ExtType; // sign extend - defm : MemOpi_m5ExtType; // any extend +let Predicates = [UseMEMOP] in { + defm: MemOpi_m5ExtType; // zero extend + defm: MemOpi_m5ExtType; // sign extend + defm: MemOpi_m5ExtType; // any extend // Word - defm : MemOpi_m5Pats ; + defm: MemOpi_m5Pats ; } //===----------------------------------------------------------------------===// @@ -2556,52 +3060,50 @@ let Predicates = [HasV4T, UseMEMOP] in { //===----------------------------------------------------------------------===// multiclass MemOpi_bitPats { + PatLeaf extPred, SDNodeXForm xformFunc, InstHexagon MI, + SDNode OpNode> { // mem[bhw](Rs+#u6:[012]) = [clrbit|setbit](#U5) let AddedComplexity = 250 in - def : Pat<(stOp (OpNode (ldOp (add IntRegs:$base, extPred:$offset)), - immPred:$bitend), - (add IntRegs:$base, extPred:$offset)), - (MI IntRegs:$base, extPred:$offset, (xformFunc immPred:$bitend))>; + def: Pat<(stOp (OpNode (ldOp (add IntRegs:$base, extPred:$offset)), + immPred:$bitend), + (add IntRegs:$base, extPred:$offset)), + (MI IntRegs:$base, extPred:$offset, (xformFunc immPred:$bitend))>; // mem[bhw](Rs+#0) = [clrbit|setbit](#U5) let AddedComplexity = 225 in - def : Pat <(stOp (OpNode (ldOp (addrPred IntRegs:$addr, extPred:$offset)), - immPred:$bitend), - (addrPred (i32 IntRegs:$addr), extPred:$offset)), - (MI IntRegs:$addr, extPred:$offset, (xformFunc immPred:$bitend))>; + def: Pat<(stOp (OpNode (ldOp IntRegs:$addr), immPred:$bitend), IntRegs:$addr), + (MI IntRegs:$addr, 0, (xformFunc immPred:$bitend))>; } -multiclass MemOpi_bitExtType { +multiclass MemOpi_bitExtType { // Byte - clrbit - defm : MemOpi_bitPats; + defm: MemOpi_bitPats; // Byte - setbit - defm : MemOpi_bitPats; + defm: MemOpi_bitPats; // Half Word - clrbit - defm : MemOpi_bitPats; + defm: MemOpi_bitPats; // Half Word - setbit - defm : MemOpi_bitPats; + defm: MemOpi_bitPats; } -let Predicates = [HasV4T, UseMEMOP] in { +let Predicates = [UseMEMOP] in { // mem[bh](Rs+#0) = [clrbit|setbit](#U5) // mem[bh](Rs+#u6:[01]) = [clrbit|setbit](#U5) - defm : MemOpi_bitExtType; // zero extend - defm : MemOpi_bitExtType; // sign extend - defm : MemOpi_bitExtType; // any extend + defm: MemOpi_bitExtType; // zero extend + defm: MemOpi_bitExtType; // sign extend + defm: MemOpi_bitExtType; // any extend // memw(Rs+#0) = [clrbit|setbit](#U5) // memw(Rs+#u6:2) = [clrbit|setbit](#U5) - defm : MemOpi_bitPats; - defm : MemOpi_bitPats; + defm: MemOpi_bitPats; + defm: MemOpi_bitPats; } //===----------------------------------------------------------------------===// @@ -2611,54 +3113,51 @@ let Predicates = [HasV4T, UseMEMOP] in { // mem[bhw](Rs+#U6:[012]) [+-&|]= Rt //===----------------------------------------------------------------------===// -multiclass MemOpr_Pats { +multiclass MemOpr_Pats { let AddedComplexity = 141 in // mem[bhw](Rs+#0) [+-&|]= Rt - def : Pat <(stOp (OpNode (ldOp (addrPred IntRegs:$addr, extPred:$offset)), - (i32 IntRegs:$addend)), - (addrPred (i32 IntRegs:$addr), extPred:$offset)), - (MI IntRegs:$addr, extPred:$offset, (i32 IntRegs:$addend) )>; + def: Pat<(stOp (OpNode (ldOp IntRegs:$addr), (i32 IntRegs:$addend)), + IntRegs:$addr), + (MI IntRegs:$addr, 0, (i32 IntRegs:$addend))>; // mem[bhw](Rs+#U6:[012]) [+-&|]= Rt let AddedComplexity = 150 in - def : Pat <(stOp (OpNode (ldOp (add IntRegs:$base, extPred:$offset)), - (i32 IntRegs:$orend)), - (add IntRegs:$base, extPred:$offset)), - (MI IntRegs:$base, extPred:$offset, (i32 IntRegs:$orend) )>; + def: Pat<(stOp (OpNode (ldOp (add IntRegs:$base, extPred:$offset)), + (i32 IntRegs:$orend)), + (add IntRegs:$base, extPred:$offset)), + (MI IntRegs:$base, extPred:$offset, (i32 IntRegs:$orend))>; } -multiclass MemOPr_ALUOp { - - defm : MemOpr_Pats ; - defm : MemOpr_Pats ; - defm : MemOpr_Pats ; - defm : MemOpr_Pats ; + InstHexagon andMI, InstHexagon orMI> { + defm: MemOpr_Pats ; + defm: MemOpr_Pats ; + defm: MemOpr_Pats ; + defm: MemOpr_Pats ; } multiclass MemOPr_ExtType { // Half Word - defm : MemOPr_ALUOp ; + defm: MemOPr_ALUOp ; // Byte - defm : MemOPr_ALUOp ; + defm: MemOPr_ALUOp ; } // Define 'def Pats' for MemOps with register addend. -let Predicates = [HasV4T, UseMEMOP] in { +let Predicates = [UseMEMOP] in { // Byte, Half Word - defm : MemOPr_ExtType; // zero extend - defm : MemOPr_ExtType; // sign extend - defm : MemOPr_ExtType; // any extend + defm: MemOPr_ExtType; // zero extend + defm: MemOPr_ExtType; // sign extend + defm: MemOPr_ExtType; // any extend // Word - defm : MemOPr_ALUOp ; + defm: MemOPr_ALUOp ; } //===----------------------------------------------------------------------===// @@ -2676,311 +3175,41 @@ let Predicates = [HasV4T, UseMEMOP] in { // incorrect code for negative numbers. // Pd=cmpb.eq(Rs,#u8) -let isCompare = 1, isExtendable = 1, opExtendable = 2, hasSideEffects = 0, - validSubTargets = HasV4SubT in -class CMP_NOT_REG_IMM op, Operand ImmOp, - list Pattern> - : ALU32Inst <(outs PredRegs:$dst), (ins IntRegs:$src1, ImmOp:$src2), - "$dst = !cmp."#OpName#"($src1, #$src2)", - Pattern, - "", ALU32_2op_tc_2early_SLOT0123> { - bits<2> dst; - bits<5> src1; - bits<10> src2; +// p=!cmp.eq(r1,#s10) +def C4_cmpneqi : T_CMP <"cmp.eq", 0b00, 1, s10Ext>; +def C4_cmpltei : T_CMP <"cmp.gt", 0b01, 1, s10Ext>; +def C4_cmplteui : T_CMP <"cmp.gtu", 0b10, 1, u9Ext>; - let IClass = 0b0111; - let Inst{27-24} = 0b0101; - let Inst{23-22} = op; - let Inst{20-16} = src1; - let Inst{21} = !if (!eq(OpName, "gtu"), 0b0, src2{9}); - let Inst{13-5} = src2{8-0}; - let Inst{4-2} = 0b100; - let Inst{1-0} = dst; -} - -let opExtentBits = 10, isExtentSigned = 1 in { -def C4_cmpneqi : CMP_NOT_REG_IMM <"eq", 0b00, s10Ext, [(set (i1 PredRegs:$dst), - (setne (i32 IntRegs:$src1), s10ExtPred:$src2))]>; - -def C4_cmpltei : CMP_NOT_REG_IMM <"gt", 0b01, s10Ext, [(set (i1 PredRegs:$dst), - (not (setgt (i32 IntRegs:$src1), s10ExtPred:$src2)))]>; - -} -let opExtentBits = 9 in -def C4_cmplteui : CMP_NOT_REG_IMM <"gtu", 0b10, u9Ext, [(set (i1 PredRegs:$dst), - (not (setugt (i32 IntRegs:$src1), u9ExtPred:$src2)))]>; - - - -// p=!cmp.eq(r1,r2) -let isCompare = 1, validSubTargets = HasV4SubT in -def CMPnotEQ_rr : ALU32_rr<(outs PredRegs:$dst), - (ins IntRegs:$src1, IntRegs:$src2), - "$dst = !cmp.eq($src1, $src2)", - [(set (i1 PredRegs:$dst), - (setne (i32 IntRegs:$src1), (i32 IntRegs:$src2)))]>, - Requires<[HasV4T]>; - -// p=!cmp.gt(r1,r2) -let isCompare = 1, validSubTargets = HasV4SubT in -def CMPnotGT_rr : ALU32_rr<(outs PredRegs:$dst), - (ins IntRegs:$src1, IntRegs:$src2), - "$dst = !cmp.gt($src1, $src2)", - [(set (i1 PredRegs:$dst), - (not (setgt (i32 IntRegs:$src1), (i32 IntRegs:$src2))))]>, - Requires<[HasV4T]>; - - -// p=!cmp.gtu(r1,r2) -let isCompare = 1, validSubTargets = HasV4SubT in -def CMPnotGTU_rr : ALU32_rr<(outs PredRegs:$dst), - (ins IntRegs:$src1, IntRegs:$src2), - "$dst = !cmp.gtu($src1, $src2)", - [(set (i1 PredRegs:$dst), - (not (setugt (i32 IntRegs:$src1), (i32 IntRegs:$src2))))]>, - Requires<[HasV4T]>; - -let isCompare = 1, validSubTargets = HasV4SubT in -def CMPbEQri_V4 : MInst<(outs PredRegs:$dst), - (ins IntRegs:$src1, u8Imm:$src2), - "$dst = cmpb.eq($src1, #$src2)", - [(set (i1 PredRegs:$dst), - (seteq (and (i32 IntRegs:$src1), 255), u8ImmPred:$src2))]>, - Requires<[HasV4T]>; - -def : Pat <(brcond (i1 (setne (and (i32 IntRegs:$src1), 255), u8ImmPred:$src2)), - bb:$offset), - (J2_jumpf (CMPbEQri_V4 (i32 IntRegs:$src1), u8ImmPred:$src2), - bb:$offset)>, - Requires<[HasV4T]>; - -// Pd=cmpb.eq(Rs,Rt) -let isCompare = 1, validSubTargets = HasV4SubT in -def CMPbEQrr_ubub_V4 : MInst<(outs PredRegs:$dst), - (ins IntRegs:$src1, IntRegs:$src2), - "$dst = cmpb.eq($src1, $src2)", - [(set (i1 PredRegs:$dst), - (seteq (and (xor (i32 IntRegs:$src1), - (i32 IntRegs:$src2)), 255), 0))]>, - Requires<[HasV4T]>; - -// Pd=cmpb.eq(Rs,Rt) -let isCompare = 1, validSubTargets = HasV4SubT in -def CMPbEQrr_sbsb_V4 : MInst<(outs PredRegs:$dst), - (ins IntRegs:$src1, IntRegs:$src2), - "$dst = cmpb.eq($src1, $src2)", - [(set (i1 PredRegs:$dst), - (seteq (shl (i32 IntRegs:$src1), (i32 24)), - (shl (i32 IntRegs:$src2), (i32 24))))]>, - Requires<[HasV4T]>; - -// Pd=cmpb.gt(Rs,Rt) -let isCompare = 1, validSubTargets = HasV4SubT in -def CMPbGTrr_V4 : MInst<(outs PredRegs:$dst), - (ins IntRegs:$src1, IntRegs:$src2), - "$dst = cmpb.gt($src1, $src2)", - [(set (i1 PredRegs:$dst), - (setgt (shl (i32 IntRegs:$src1), (i32 24)), - (shl (i32 IntRegs:$src2), (i32 24))))]>, - Requires<[HasV4T]>; - -// Pd=cmpb.gtu(Rs,#u7) -let isExtendable = 1, opExtendable = 2, isExtentSigned = 0, opExtentBits = 7, -isCompare = 1, validSubTargets = HasV4SubT, CextOpcode = "CMPbGTU", InputType = "imm" in -def CMPbGTUri_V4 : MInst<(outs PredRegs:$dst), - (ins IntRegs:$src1, u7Ext:$src2), - "$dst = cmpb.gtu($src1, #$src2)", - [(set (i1 PredRegs:$dst), (setugt (and (i32 IntRegs:$src1), 255), - u7ExtPred:$src2))]>, - Requires<[HasV4T]>, ImmRegRel; +def : T_CMP_pat ; +def : T_CMP_pat ; +def : T_CMP_pat ; + +// rs <= rt -> !(rs > rt). +/* +def: Pat<(i1 (setle (i32 IntRegs:$src1), s32ImmPred:$src2)), + (C2_not (C2_cmpgti IntRegs:$src1, s32ImmPred:$src2))>; +// (C4_cmpltei IntRegs:$src1, s32ImmPred:$src2)>; +*/ +// Map cmplt(Rs, Imm) -> !cmpgt(Rs, Imm-1). +def: Pat<(i1 (setlt (i32 IntRegs:$src1), s32ImmPred:$src2)), + (C4_cmpltei IntRegs:$src1, (DEC_CONST_SIGNED s32ImmPred:$src2))>; + +// rs != rt -> !(rs == rt). +def: Pat<(i1 (setne (i32 IntRegs:$src1), s32ImmPred:$src2)), + (C4_cmpneqi IntRegs:$src1, s32ImmPred:$src2)>; // SDNode for converting immediate C to C-1. def DEC_CONST_BYTE : SDNodeXFormgetSExtValue(); - return XformU7ToU7M1Imm(imm); + return XformU7ToU7M1Imm(imm, SDLoc(N)); }]>; -// For the sequence -// zext( seteq ( and(Rs, 255), u8)) -// Generate -// Pd=cmpb.eq(Rs, #u8) -// if (Pd.new) Rd=#1 -// if (!Pd.new) Rd=#0 -def : Pat <(i32 (zext (i1 (seteq (i32 (and (i32 IntRegs:$Rs), 255)), - u8ExtPred:$u8)))), - (i32 (TFR_condset_ii (i1 (CMPbEQri_V4 (i32 IntRegs:$Rs), - (u8ExtPred:$u8))), - 1, 0))>, - Requires<[HasV4T]>; - -// For the sequence -// zext( setne ( and(Rs, 255), u8)) -// Generate -// Pd=cmpb.eq(Rs, #u8) -// if (Pd.new) Rd=#0 -// if (!Pd.new) Rd=#1 -def : Pat <(i32 (zext (i1 (setne (i32 (and (i32 IntRegs:$Rs), 255)), - u8ExtPred:$u8)))), - (i32 (TFR_condset_ii (i1 (CMPbEQri_V4 (i32 IntRegs:$Rs), - (u8ExtPred:$u8))), - 0, 1))>, - Requires<[HasV4T]>; - -// For the sequence -// zext( seteq (Rs, and(Rt, 255))) -// Generate -// Pd=cmpb.eq(Rs, Rt) -// if (Pd.new) Rd=#1 -// if (!Pd.new) Rd=#0 -def : Pat <(i32 (zext (i1 (seteq (i32 IntRegs:$Rt), - (i32 (and (i32 IntRegs:$Rs), 255)))))), - (i32 (TFR_condset_ii (i1 (CMPbEQrr_ubub_V4 (i32 IntRegs:$Rs), - (i32 IntRegs:$Rt))), - 1, 0))>, - Requires<[HasV4T]>; - -// For the sequence -// zext( setne (Rs, and(Rt, 255))) -// Generate -// Pd=cmpb.eq(Rs, Rt) -// if (Pd.new) Rd=#0 -// if (!Pd.new) Rd=#1 -def : Pat <(i32 (zext (i1 (setne (i32 IntRegs:$Rt), - (i32 (and (i32 IntRegs:$Rs), 255)))))), - (i32 (TFR_condset_ii (i1 (CMPbEQrr_ubub_V4 (i32 IntRegs:$Rs), - (i32 IntRegs:$Rt))), - 0, 1))>, - Requires<[HasV4T]>; - -// For the sequence -// zext( setugt ( and(Rs, 255), u8)) -// Generate -// Pd=cmpb.gtu(Rs, #u8) -// if (Pd.new) Rd=#1 -// if (!Pd.new) Rd=#0 -def : Pat <(i32 (zext (i1 (setugt (i32 (and (i32 IntRegs:$Rs), 255)), - u8ExtPred:$u8)))), - (i32 (TFR_condset_ii (i1 (CMPbGTUri_V4 (i32 IntRegs:$Rs), - (u8ExtPred:$u8))), - 1, 0))>, - Requires<[HasV4T]>; - -// For the sequence -// zext( setugt ( and(Rs, 254), u8)) -// Generate -// Pd=cmpb.gtu(Rs, #u8) -// if (Pd.new) Rd=#1 -// if (!Pd.new) Rd=#0 -def : Pat <(i32 (zext (i1 (setugt (i32 (and (i32 IntRegs:$Rs), 254)), - u8ExtPred:$u8)))), - (i32 (TFR_condset_ii (i1 (CMPbGTUri_V4 (i32 IntRegs:$Rs), - (u8ExtPred:$u8))), - 1, 0))>, - Requires<[HasV4T]>; - -// For the sequence -// zext( setult ( Rs, Rt)) -// Generate -// Pd=cmp.ltu(Rs, Rt) -// if (Pd.new) Rd=#1 -// if (!Pd.new) Rd=#0 -// cmp.ltu(Rs, Rt) -> cmp.gtu(Rt, Rs) -def : Pat <(i32 (zext (i1 (setult (i32 IntRegs:$Rs), (i32 IntRegs:$Rt))))), - (i32 (TFR_condset_ii (i1 (C2_cmpgtu (i32 IntRegs:$Rt), - (i32 IntRegs:$Rs))), - 1, 0))>, - Requires<[HasV4T]>; - -// For the sequence -// zext( setlt ( Rs, Rt)) -// Generate -// Pd=cmp.lt(Rs, Rt) -// if (Pd.new) Rd=#1 -// if (!Pd.new) Rd=#0 -// cmp.lt(Rs, Rt) -> cmp.gt(Rt, Rs) -def : Pat <(i32 (zext (i1 (setlt (i32 IntRegs:$Rs), (i32 IntRegs:$Rt))))), - (i32 (TFR_condset_ii (i1 (C2_cmpgt (i32 IntRegs:$Rt), - (i32 IntRegs:$Rs))), - 1, 0))>, - Requires<[HasV4T]>; - -// For the sequence -// zext( setugt ( Rs, Rt)) -// Generate -// Pd=cmp.gtu(Rs, Rt) -// if (Pd.new) Rd=#1 -// if (!Pd.new) Rd=#0 -def : Pat <(i32 (zext (i1 (setugt (i32 IntRegs:$Rs), (i32 IntRegs:$Rt))))), - (i32 (TFR_condset_ii (i1 (C2_cmpgtu (i32 IntRegs:$Rs), - (i32 IntRegs:$Rt))), - 1, 0))>, - Requires<[HasV4T]>; - -// This pattern interefers with coremark performance, not implementing at this -// time. -// For the sequence -// zext( setgt ( Rs, Rt)) -// Generate -// Pd=cmp.gt(Rs, Rt) -// if (Pd.new) Rd=#1 -// if (!Pd.new) Rd=#0 - -// For the sequence -// zext( setuge ( Rs, Rt)) -// Generate -// Pd=cmp.ltu(Rs, Rt) -// if (Pd.new) Rd=#0 -// if (!Pd.new) Rd=#1 -// cmp.ltu(Rs, Rt) -> cmp.gtu(Rt, Rs) -def : Pat <(i32 (zext (i1 (setuge (i32 IntRegs:$Rs), (i32 IntRegs:$Rt))))), - (i32 (TFR_condset_ii (i1 (C2_cmpgtu (i32 IntRegs:$Rt), - (i32 IntRegs:$Rs))), - 0, 1))>, - Requires<[HasV4T]>; - -// For the sequence -// zext( setge ( Rs, Rt)) -// Generate -// Pd=cmp.lt(Rs, Rt) -// if (Pd.new) Rd=#0 -// if (!Pd.new) Rd=#1 -// cmp.lt(Rs, Rt) -> cmp.gt(Rt, Rs) -def : Pat <(i32 (zext (i1 (setge (i32 IntRegs:$Rs), (i32 IntRegs:$Rt))))), - (i32 (TFR_condset_ii (i1 (C2_cmpgt (i32 IntRegs:$Rt), - (i32 IntRegs:$Rs))), - 0, 1))>, - Requires<[HasV4T]>; - -// For the sequence -// zext( setule ( Rs, Rt)) -// Generate -// Pd=cmp.gtu(Rs, Rt) -// if (Pd.new) Rd=#0 -// if (!Pd.new) Rd=#1 -def : Pat <(i32 (zext (i1 (setule (i32 IntRegs:$Rs), (i32 IntRegs:$Rt))))), - (i32 (TFR_condset_ii (i1 (C2_cmpgtu (i32 IntRegs:$Rs), - (i32 IntRegs:$Rt))), - 0, 1))>, - Requires<[HasV4T]>; - -// For the sequence -// zext( setle ( Rs, Rt)) -// Generate -// Pd=cmp.gt(Rs, Rt) -// if (Pd.new) Rd=#0 -// if (!Pd.new) Rd=#1 -def : Pat <(i32 (zext (i1 (setle (i32 IntRegs:$Rs), (i32 IntRegs:$Rt))))), - (i32 (TFR_condset_ii (i1 (C2_cmpgt (i32 IntRegs:$Rs), - (i32 IntRegs:$Rt))), - 0, 1))>, - Requires<[HasV4T]>; - // For the sequence // zext( setult ( and(Rs, 255), u8)) // Use the isdigit transformation below -// Generate code of the form 'mux_ii(cmpbgtu(Rdd, C-1),0,1)' +// Generate code of the form 'C2_muxii(cmpbgtui(Rdd, C-1),0,1)' // for C code of the form r = ((c>='0') & (c<='9')) ? 1 : 0;. // The isdigit transformation relies on two 'clever' aspects: // 1) The data type is unsigned which allows us to eliminate a zero test after @@ -2993,130 +3222,11 @@ def : Pat <(i32 (zext (i1 (setle (i32 IntRegs:$Rs), (i32 IntRegs:$Rt))))), // The code is transformed upstream of llvm into // retval = (c-48) < 10 ? 1 : 0; let AddedComplexity = 139 in -def : Pat <(i32 (zext (i1 (setult (i32 (and (i32 IntRegs:$src1), 255)), - u7StrictPosImmPred:$src2)))), - (i32 (C2_muxii (i1 (CMPbGTUri_V4 (i32 IntRegs:$src1), - (DEC_CONST_BYTE u7StrictPosImmPred:$src2))), - 0, 1))>, - Requires<[HasV4T]>; - -// Pd=cmpb.gtu(Rs,Rt) -let isCompare = 1, validSubTargets = HasV4SubT, CextOpcode = "CMPbGTU", -InputType = "reg" in -def CMPbGTUrr_V4 : MInst<(outs PredRegs:$dst), - (ins IntRegs:$src1, IntRegs:$src2), - "$dst = cmpb.gtu($src1, $src2)", - [(set (i1 PredRegs:$dst), (setugt (and (i32 IntRegs:$src1), 255), - (and (i32 IntRegs:$src2), 255)))]>, - Requires<[HasV4T]>, ImmRegRel; - -// Following instruction is not being extended as it results into the incorrect -// code for negative numbers. - -// Signed half compare(.eq) ri. -// Pd=cmph.eq(Rs,#s8) -let isCompare = 1, validSubTargets = HasV4SubT in -def CMPhEQri_V4 : MInst<(outs PredRegs:$dst), - (ins IntRegs:$src1, s8Imm:$src2), - "$dst = cmph.eq($src1, #$src2)", - [(set (i1 PredRegs:$dst), (seteq (and (i32 IntRegs:$src1), 65535), - s8ImmPred:$src2))]>, - Requires<[HasV4T]>; - -// Signed half compare(.eq) rr. -// Case 1: xor + and, then compare: -// r0=xor(r0,r1) -// r0=and(r0,#0xffff) -// p0=cmp.eq(r0,#0) -// Pd=cmph.eq(Rs,Rt) -let isCompare = 1, validSubTargets = HasV4SubT in -def CMPhEQrr_xor_V4 : MInst<(outs PredRegs:$dst), - (ins IntRegs:$src1, IntRegs:$src2), - "$dst = cmph.eq($src1, $src2)", - [(set (i1 PredRegs:$dst), (seteq (and (xor (i32 IntRegs:$src1), - (i32 IntRegs:$src2)), - 65535), 0))]>, - Requires<[HasV4T]>; - -// Signed half compare(.eq) rr. -// Case 2: shift left 16 bits then compare: -// r0=asl(r0,16) -// r1=asl(r1,16) -// p0=cmp.eq(r0,r1) -// Pd=cmph.eq(Rs,Rt) -let isCompare = 1, validSubTargets = HasV4SubT in -def CMPhEQrr_shl_V4 : MInst<(outs PredRegs:$dst), - (ins IntRegs:$src1, IntRegs:$src2), - "$dst = cmph.eq($src1, $src2)", - [(set (i1 PredRegs:$dst), - (seteq (shl (i32 IntRegs:$src1), (i32 16)), - (shl (i32 IntRegs:$src2), (i32 16))))]>, - Requires<[HasV4T]>; - -/* Incorrect Pattern -- immediate should be right shifted before being -used in the cmph.gt instruction. -// Signed half compare(.gt) ri. -// Pd=cmph.gt(Rs,#s8) - -let isExtendable = 1, opExtendable = 2, isExtentSigned = 1, opExtentBits = 8, -isCompare = 1, validSubTargets = HasV4SubT in -def CMPhGTri_V4 : MInst<(outs PredRegs:$dst), - (ins IntRegs:$src1, s8Ext:$src2), - "$dst = cmph.gt($src1, #$src2)", - [(set (i1 PredRegs:$dst), - (setgt (shl (i32 IntRegs:$src1), (i32 16)), - s8ExtPred:$src2))]>, - Requires<[HasV4T]>; -*/ - -// Signed half compare(.gt) rr. -// Pd=cmph.gt(Rs,Rt) -let isCompare = 1, validSubTargets = HasV4SubT in -def CMPhGTrr_shl_V4 : MInst<(outs PredRegs:$dst), - (ins IntRegs:$src1, IntRegs:$src2), - "$dst = cmph.gt($src1, $src2)", - [(set (i1 PredRegs:$dst), - (setgt (shl (i32 IntRegs:$src1), (i32 16)), - (shl (i32 IntRegs:$src2), (i32 16))))]>, - Requires<[HasV4T]>; - -// Unsigned half compare rr (.gtu). -// Pd=cmph.gtu(Rs,Rt) -let isCompare = 1, validSubTargets = HasV4SubT, CextOpcode = "CMPhGTU", -InputType = "reg" in -def CMPhGTUrr_V4 : MInst<(outs PredRegs:$dst), - (ins IntRegs:$src1, IntRegs:$src2), - "$dst = cmph.gtu($src1, $src2)", - [(set (i1 PredRegs:$dst), - (setugt (and (i32 IntRegs:$src1), 65535), - (and (i32 IntRegs:$src2), 65535)))]>, - Requires<[HasV4T]>, ImmRegRel; - -// Unsigned half compare ri (.gtu). -// Pd=cmph.gtu(Rs,#u7) -let isExtendable = 1, opExtendable = 2, isExtentSigned = 0, opExtentBits = 7, -isCompare = 1, validSubTargets = HasV4SubT, CextOpcode = "CMPhGTU", -InputType = "imm" in -def CMPhGTUri_V4 : MInst<(outs PredRegs:$dst), - (ins IntRegs:$src1, u7Ext:$src2), - "$dst = cmph.gtu($src1, #$src2)", - [(set (i1 PredRegs:$dst), (setugt (and (i32 IntRegs:$src1), 65535), - u7ExtPred:$src2))]>, - Requires<[HasV4T]>, ImmRegRel; - -let validSubTargets = HasV4SubT in -def NTSTBIT_rr : SInst<(outs PredRegs:$dst), (ins IntRegs:$src1, IntRegs:$src2), - "$dst = !tstbit($src1, $src2)", - [(set (i1 PredRegs:$dst), - (seteq (and (shl 1, (i32 IntRegs:$src2)), (i32 IntRegs:$src1)), 0))]>, - Requires<[HasV4T]>; - -let validSubTargets = HasV4SubT in -def NTSTBIT_ri : SInst<(outs PredRegs:$dst), (ins IntRegs:$src1, u5Imm:$src2), - "$dst = !tstbit($src1, $src2)", - [(set (i1 PredRegs:$dst), - (seteq (and (shl 1, u5ImmPred:$src2), (i32 IntRegs:$src1)), 0))]>, - Requires<[HasV4T]>; +def: Pat<(i32 (zext (i1 (setult (i32 (and (i32 IntRegs:$src1), 255)), + u7StrictPosImmPred:$src2)))), + (C2_muxii (A4_cmpbgtui IntRegs:$src1, + (DEC_CONST_BYTE u7StrictPosImmPred:$src2)), + 0, 1)>; //===----------------------------------------------------------------------===// // XTYPE/PRED - @@ -3173,40 +3283,23 @@ multiclass LD_MISC_L4_RETURN { } let isReturn = 1, isTerminator = 1, - Defs = [R29, R30, R31, PC], Uses = [R30], hasSideEffects = 0, - validSubTargets = HasV4SubT, isCodeGenOnly = 0 in + Defs = [R29, R30, R31, PC], Uses = [R30], hasSideEffects = 0 in defm L4_return: LD_MISC_L4_RETURN <"dealloc_return">, PredNewRel; // Restore registers and dealloc return function call. let isCall = 1, isBarrier = 1, isReturn = 1, isTerminator = 1, - Defs = [R29, R30, R31, PC] in { -let validSubTargets = HasV4SubT in - def RESTORE_DEALLOC_RET_JMP_V4 : JInst<(outs), - (ins calltarget:$dst), - "jump $dst", - []>, - Requires<[HasV4T]>; + Defs = [R29, R30, R31, PC], isPredicable = 0, isAsmParserOnly = 1 in { + def RESTORE_DEALLOC_RET_JMP_V4 : T_JMP<"">; } // Restore registers and dealloc frame before a tail call. -let isCall = 1, isBarrier = 1, - Defs = [R29, R30, R31, PC] in { -let validSubTargets = HasV4SubT in - def RESTORE_DEALLOC_BEFORE_TAILCALL_V4 : JInst<(outs), - (ins calltarget:$dst), - "call $dst", - []>, - Requires<[HasV4T]>; +let isCall = 1, Defs = [R29, R30, R31, PC], isAsmParserOnly = 1 in { + def RESTORE_DEALLOC_BEFORE_TAILCALL_V4 : T_Call<"">, PredRel; } // Save registers function call. -let isCall = 1, isBarrier = 1, - Uses = [R29, R31] in { - def SAVE_REGISTERS_CALL_V4 : JInst<(outs), - (ins calltarget:$dst), - "call $dst // Save_calle_saved_registers", - []>, - Requires<[HasV4T]>; +let isCall = 1, Uses = [R29, R31], isAsmParserOnly = 1 in { + def SAVE_REGISTERS_CALL_V4 : T_Call<"">, PredRel; } //===----------------------------------------------------------------------===// @@ -3278,7 +3371,7 @@ class T_StoreAbs_Pred MajOp, //===----------------------------------------------------------------------===// class T_StoreAbs MajOp, bit isHalf> - : T_StoreAbsGP , + : T_StoreAbsGP , AddrModeRel { string ImmOpStr = !cast(ImmOp); let opExtentBits = !if (!eq(ImmOpStr, "u16_3Imm"), 19, @@ -3295,7 +3388,7 @@ class T_StoreAbs MajOp, bit isHalf = 0> { let CextOpcode = CextOp, BaseOpcode = CextOp#_abs in { @@ -3319,7 +3412,7 @@ multiclass ST_AbsMajOp, bit isAbs> - : NVInst_V4<(outs), (ins u0AlwaysExt:$addr, IntRegs:$src), + : NVInst_V4<(outs), (ins u32Imm:$addr, IntRegs:$src), mnemonic # !if(isAbs, "(##", "(#")#"$addr) = $src.new", [], "", V2LDST_tc_st_SLOT0> { bits<19> addr; @@ -3397,7 +3490,7 @@ class T_StoreAbs_NV MajOp> //===----------------------------------------------------------------------===// // Multiclass for new-value store instructions with absolute addressing. //===----------------------------------------------------------------------===// -let validSubTargets = HasV4SubT, addrMode = Absolute, isExtended = 1 in +let addrMode = Absolute, isExtended = 1 in multiclass ST_Abs_NV MajOp> { let CextOpcode = CextOp, BaseOpcode = CextOp#_abs in { @@ -3417,22 +3510,22 @@ multiclass ST_Abs_NV , ST_Abs_NV <"memb", "STrib", u16_0Imm, 0b00>; -let accessSize = HalfWordAccess, isCodeGenOnly = 0 in +let accessSize = HalfWordAccess in defm storerh : ST_Abs <"memh", "STrih", IntRegs, u16_1Imm, 0b01>, ST_Abs_NV <"memh", "STrih", u16_1Imm, 0b01>; -let accessSize = WordAccess, isCodeGenOnly = 0 in +let accessSize = WordAccess in defm storeri : ST_Abs <"memw", "STriw", IntRegs, u16_2Imm, 0b10>, ST_Abs_NV <"memw", "STriw", u16_2Imm, 0b10>; -let isNVStorable = 0, accessSize = DoubleWordAccess, isCodeGenOnly = 0 in +let isNVStorable = 0, accessSize = DoubleWordAccess in defm storerd : ST_Abs <"memd", "STrid", DoubleRegs, u16_3Imm, 0b11>; -let isNVStorable = 0, accessSize = HalfWordAccess, isCodeGenOnly = 0 in +let isNVStorable = 0, accessSize = HalfWordAccess in defm storerf : ST_Abs <"memh", "STrif", IntRegs, u16_1Imm, 0b01, 1>; //===----------------------------------------------------------------------===// @@ -3442,7 +3535,7 @@ defm storerf : ST_Abs <"memh", "STrif", IntRegs, u16_1Imm, 0b01, 1>; // if ([!]Pv[.new]) mem[bhwd](##global)=Rt //===----------------------------------------------------------------------===// -let validSubTargets = HasV4SubT in +let isAsmParserOnly = 1 in class T_StoreGP MajOp, bit isHalf = 0> : T_StoreAbsGP { @@ -3452,7 +3545,7 @@ class T_StoreGP MajOp, bit isHalf = 0> { // Set BaseOpcode same as absolute addressing instructions so that @@ -3483,77 +3576,44 @@ let isNVStorable = 0, accessSize = HalfWordAccess in def S2_storerfgp : T_StoreGP <"memh", "STrif", IntRegs, u16_1Imm, 0b01, 1>, PredNewRel; -let Predicates = [HasV4T], AddedComplexity = 30 in { -def : Pat<(truncstorei8 (i32 IntRegs:$src1), - (HexagonCONST32 tglobaladdr:$absaddr)), - (S2_storerbabs tglobaladdr: $absaddr, IntRegs: $src1)>; - -def : Pat<(truncstorei16 (i32 IntRegs:$src1), - (HexagonCONST32 tglobaladdr:$absaddr)), - (S2_storerhabs tglobaladdr: $absaddr, IntRegs: $src1)>; - -def : Pat<(store (i32 IntRegs:$src1), (HexagonCONST32 tglobaladdr:$absaddr)), - (S2_storeriabs tglobaladdr: $absaddr, IntRegs: $src1)>; - -def : Pat<(store (i64 DoubleRegs:$src1), - (HexagonCONST32 tglobaladdr:$absaddr)), - (S2_storerdabs tglobaladdr: $absaddr, DoubleRegs: $src1)>; +class Loada_pat + : Pat<(VT (Load Addr:$addr)), (MI Addr:$addr)>; + +class Loadam_pat + : Pat<(VT (Load Addr:$addr)), (ValueMod (MI Addr:$addr))>; + +class Storea_pat + : Pat<(Store Value:$val, Addr:$addr), (MI Addr:$addr, Value:$val)>; + +class Stoream_pat + : Pat<(Store Value:$val, Addr:$addr), + (MI Addr:$addr, (ValueMod Value:$val))>; + +def: Storea_pat, I32, addrgp, S2_storerbgp>; +def: Storea_pat, I32, addrgp, S2_storerhgp>; +def: Storea_pat, I32, addrgp, S2_storerigp>; +def: Storea_pat, I64, addrgp, S2_storerdgp>; + +let AddedComplexity = 100 in { + def: Storea_pat; + def: Storea_pat; + def: Storea_pat; + def: Storea_pat; + + // Map from "i1 = constant<-1>; memw(CONST32(#foo)) = i1" + // to "r0 = 1; memw(#foo) = r0" + let AddedComplexity = 100 in + def: Pat<(store (i1 -1), (HexagonCONST32_GP tglobaladdr:$global)), + (S2_storerbgp tglobaladdr:$global, (A2_tfrsi 1))>; } -// 64 bit atomic store -def : Pat <(atomic_store_64 (HexagonCONST32_GP tglobaladdr:$global), - (i64 DoubleRegs:$src1)), - (S2_storerdgp tglobaladdr:$global, (i64 DoubleRegs:$src1))>, - Requires<[HasV4T]>; - -// Map from store(globaladdress) -> memd(#foo) -let AddedComplexity = 100 in -def : Pat <(store (i64 DoubleRegs:$src1), - (HexagonCONST32_GP tglobaladdr:$global)), - (S2_storerdgp tglobaladdr:$global, (i64 DoubleRegs:$src1))>; - -// 8 bit atomic store -def : Pat < (atomic_store_8 (HexagonCONST32_GP tglobaladdr:$global), - (i32 IntRegs:$src1)), - (S2_storerbgp tglobaladdr:$global, (i32 IntRegs:$src1))>; - -// Map from store(globaladdress) -> memb(#foo) -let AddedComplexity = 100 in -def : Pat<(truncstorei8 (i32 IntRegs:$src1), - (HexagonCONST32_GP tglobaladdr:$global)), - (S2_storerbgp tglobaladdr:$global, (i32 IntRegs:$src1))>; - -// Map from "i1 = constant<-1>; memw(CONST32(#foo)) = i1" -// to "r0 = 1; memw(#foo) = r0" -let AddedComplexity = 100 in -def : Pat<(store (i1 -1), (HexagonCONST32_GP tglobaladdr:$global)), - (S2_storerbgp tglobaladdr:$global, (A2_tfrsi 1))>; - -def : Pat<(atomic_store_16 (HexagonCONST32_GP tglobaladdr:$global), - (i32 IntRegs:$src1)), - (S2_storerhgp tglobaladdr:$global, (i32 IntRegs:$src1))>; - -// Map from store(globaladdress) -> memh(#foo) -let AddedComplexity = 100 in -def : Pat<(truncstorei16 (i32 IntRegs:$src1), - (HexagonCONST32_GP tglobaladdr:$global)), - (S2_storerhgp tglobaladdr:$global, (i32 IntRegs:$src1))>; - -// 32 bit atomic store -def : Pat<(atomic_store_32 (HexagonCONST32_GP tglobaladdr:$global), - (i32 IntRegs:$src1)), - (S2_storerigp tglobaladdr:$global, (i32 IntRegs:$src1))>; - -// Map from store(globaladdress) -> memw(#foo) -let AddedComplexity = 100 in -def : Pat<(store (i32 IntRegs:$src1), (HexagonCONST32_GP tglobaladdr:$global)), - (S2_storerigp tglobaladdr:$global, (i32 IntRegs:$src1))>; - //===----------------------------------------------------------------------===// // Template class for non predicated load instructions with // absolute addressing mode. //===----------------------------------------------------------------------===// -let isPredicable = 1, hasSideEffects = 0, validSubTargets = HasV4SubT in +let isPredicable = 1, hasSideEffects = 0 in class T_LoadAbsGP MajOp, Operand AddrOp, bit isAbs> : LDInst <(outs RC:$dst), (ins AddrOp:$addr), @@ -3582,7 +3642,7 @@ class T_LoadAbsGP MajOp> - : T_LoadAbsGP , AddrModeRel { + : T_LoadAbsGP , AddrModeRel { string ImmOpStr = !cast(ImmOp); let opExtentBits = !if (!eq(ImmOpStr, "u16_3Imm"), 19, @@ -3595,11 +3655,12 @@ class T_LoadAbs MajOp, bit isPredNot, bit isPredNew> : LDInst <(outs RC:$dst), (ins PredRegs:$src1, u6Ext:$absaddr), @@ -3611,6 +3672,7 @@ class T_LoadAbs_Pred MajOp, let isPredicatedNew = isPredNew; let isPredicatedFalse = isPredNot; + let hasNewValue = !if (!eq(!cast(RC), "DoubleRegs"), 0, 1); let IClass = 0b1001; @@ -3649,20 +3711,20 @@ multiclass LD_Abs; defm loadrub : LD_Abs<"memub", "LDriub", IntRegs, u16_0Imm, 0b001>; } -let accessSize = HalfWordAccess, hasNewValue = 1, isCodeGenOnly = 0 in { +let accessSize = HalfWordAccess, hasNewValue = 1 in { defm loadrh : LD_Abs<"memh", "LDrih", IntRegs, u16_1Imm, 0b010>; defm loadruh : LD_Abs<"memuh", "LDriuh", IntRegs, u16_1Imm, 0b011>; } -let accessSize = WordAccess, hasNewValue = 1, isCodeGenOnly = 0 in +let accessSize = WordAccess, hasNewValue = 1 in defm loadri : LD_Abs<"memw", "LDriw", IntRegs, u16_2Imm, 0b100>; -let accessSize = DoubleWordAccess, isCodeGenOnly = 0 in +let accessSize = DoubleWordAccess in defm loadrd : LD_Abs<"memd", "LDrid", DoubleRegs, u16_3Imm, 0b110>; //===----------------------------------------------------------------------===// @@ -3672,6 +3734,7 @@ defm loadrd : LD_Abs<"memd", "LDrid", DoubleRegs, u16_3Imm, 0b110>; // if ([!]Pv[.new]) Rx=mem[bhwd](##global) //===----------------------------------------------------------------------===// +let isAsmParserOnly = 1 in class T_LoadGP MajOp> : T_LoadAbsGP , PredNewRel { @@ -3694,439 +3757,175 @@ def L2_loadrigp : T_LoadGP<"memw", "LDriw", IntRegs, u16_2Imm, 0b100>; let accessSize = DoubleWordAccess in def L2_loadrdgp : T_LoadGP<"memd", "LDrid", DoubleRegs, u16_3Imm, 0b110>; -let Predicates = [HasV4T], AddedComplexity = 30 in { -def : Pat<(i32 (load (HexagonCONST32 tglobaladdr:$absaddr))), - (L4_loadri_abs tglobaladdr: $absaddr)>; - -def : Pat<(i32 (sextloadi8 (HexagonCONST32 tglobaladdr:$absaddr))), - (L4_loadrb_abs tglobaladdr:$absaddr)>; - -def : Pat<(i32 (zextloadi8 (HexagonCONST32 tglobaladdr:$absaddr))), - (L4_loadrub_abs tglobaladdr:$absaddr)>; - -def : Pat<(i32 (sextloadi16 (HexagonCONST32 tglobaladdr:$absaddr))), - (L4_loadrh_abs tglobaladdr:$absaddr)>; - -def : Pat<(i32 (zextloadi16 (HexagonCONST32 tglobaladdr:$absaddr))), - (L4_loadruh_abs tglobaladdr:$absaddr)>; -} - -def : Pat <(atomic_load_64 (HexagonCONST32_GP tglobaladdr:$global)), - (i64 (L2_loadrdgp tglobaladdr:$global))>; - -def : Pat <(atomic_load_32 (HexagonCONST32_GP tglobaladdr:$global)), - (i32 (L2_loadrigp tglobaladdr:$global))>; - -def : Pat <(atomic_load_16 (HexagonCONST32_GP tglobaladdr:$global)), - (i32 (L2_loadruhgp tglobaladdr:$global))>; - -def : Pat <(atomic_load_8 (HexagonCONST32_GP tglobaladdr:$global)), - (i32 (L2_loadrubgp tglobaladdr:$global))>; - -// Map from load(globaladdress) -> memw(#foo + 0) -let AddedComplexity = 100 in -def : Pat <(i64 (load (HexagonCONST32_GP tglobaladdr:$global))), - (i64 (L2_loadrdgp tglobaladdr:$global))>; +def: Loada_pat; +def: Loada_pat; +def: Loada_pat; +def: Loada_pat; // Map from Pd = load(globaladdress) -> Rd = memb(globaladdress), Pd = Rd -let AddedComplexity = 100 in -def : Pat <(i1 (load (HexagonCONST32_GP tglobaladdr:$global))), - (i1 (C2_tfrrp (i32 (L2_loadrbgp tglobaladdr:$global))))>; +def: Loadam_pat; +def: Loadam_pat; + +def: Stoream_pat; +def: Stoream_pat; + +// Map from load(globaladdress) -> mem[u][bhwd](#foo) +class LoadGP_pats + : Pat <(VT (ldOp (HexagonCONST32_GP tglobaladdr:$global))), + (VT (MI tglobaladdr:$global))>; + +let AddedComplexity = 100 in { + def: LoadGP_pats ; + def: LoadGP_pats ; + def: LoadGP_pats ; + def: LoadGP_pats ; + def: LoadGP_pats ; + def: LoadGP_pats ; + def: LoadGP_pats ; + def: LoadGP_pats ; +} // When the Interprocedural Global Variable optimizer realizes that a certain // global variable takes only two constant values, it shrinks the global to // a boolean. Catch those loads here in the following 3 patterns. -let AddedComplexity = 100 in -def : Pat <(i32 (extloadi1 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (L2_loadrbgp tglobaladdr:$global))>; +let AddedComplexity = 100 in { + def: LoadGP_pats ; + def: LoadGP_pats ; +} -let AddedComplexity = 100 in -def : Pat <(i32 (sextloadi1 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (L2_loadrbgp tglobaladdr:$global))>; +// Transfer global address into a register +def: Pat<(HexagonCONST32 tglobaladdr:$Rs), (A2_tfrsi s16Ext:$Rs)>; +def: Pat<(HexagonCONST32_GP tblockaddress:$Rs), (A2_tfrsi s16Ext:$Rs)>; +def: Pat<(HexagonCONST32_GP tglobaladdr:$Rs), (A2_tfrsi s16Ext:$Rs)>; + +let AddedComplexity = 30 in { + def: Storea_pat; + def: Storea_pat; + def: Storea_pat; +} -// Map from load(globaladdress) -> memb(#foo) -let AddedComplexity = 100 in -def : Pat <(i32 (extloadi8 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (L2_loadrbgp tglobaladdr:$global))>; +let AddedComplexity = 30 in { + def: Loada_pat; + def: Loada_pat; + def: Loada_pat; + def: Loada_pat; + def: Loada_pat; +} -// Map from load(globaladdress) -> memb(#foo) +// Indexed store word - global address. +// memw(Rs+#u6:2)=#S8 let AddedComplexity = 100 in -def : Pat <(i32 (sextloadi8 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (L2_loadrbgp tglobaladdr:$global))>; +def: Storex_add_pat; -let AddedComplexity = 100 in -def : Pat <(i32 (zextloadi1 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (L2_loadrubgp tglobaladdr:$global))>; +// Load from a global address that has only one use in the current basic block. +let AddedComplexity = 100 in { + def: Loada_pat; + def: Loada_pat; + def: Loada_pat; -// Map from load(globaladdress) -> memub(#foo) -let AddedComplexity = 100 in -def : Pat <(i32 (zextloadi8 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (L2_loadrubgp tglobaladdr:$global))>; + def: Loada_pat; + def: Loada_pat; + def: Loada_pat; -// Map from load(globaladdress) -> memh(#foo) -let AddedComplexity = 100 in -def : Pat <(i32 (extloadi16 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (L2_loadrhgp tglobaladdr:$global))>; + def: Loada_pat; + def: Loada_pat; +} -// Map from load(globaladdress) -> memh(#foo) -let AddedComplexity = 100 in -def : Pat <(i32 (sextloadi16 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (L2_loadrhgp tglobaladdr:$global))>; +// Store to a global address that has only one use in the current basic block. +let AddedComplexity = 100 in { + def: Storea_pat; + def: Storea_pat; + def: Storea_pat; + def: Storea_pat; -// Map from load(globaladdress) -> memuh(#foo) -let AddedComplexity = 100 in -def : Pat <(i32 (zextloadi16 (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (L2_loadruhgp tglobaladdr:$global))>; + def: Stoream_pat; +} -// Map from load(globaladdress) -> memw(#foo) +// Map from Pd = load(globaladdress) -> Rd = memb(globaladdress), Pd = Rd let AddedComplexity = 100 in -def : Pat <(i32 (load (HexagonCONST32_GP tglobaladdr:$global))), - (i32 (L2_loadrigp tglobaladdr:$global))>; - +def : Pat <(i1 (load (HexagonCONST32_GP tglobaladdr:$global))), + (i1 (C2_tfrrp (i32 (L2_loadrbgp tglobaladdr:$global))))>; // Transfer global address into a register let isExtended = 1, opExtendable = 1, AddedComplexity=50, isMoveImm = 1, -isAsCheapAsAMove = 1, isReMaterializable = 1, validSubTargets = HasV4SubT in +isAsCheapAsAMove = 1, isReMaterializable = 1, isCodeGenOnly = 1 in def TFRI_V4 : ALU32_ri<(outs IntRegs:$dst), (ins s16Ext:$src1), "$dst = #$src1", - [(set IntRegs:$dst, (HexagonCONST32 tglobaladdr:$src1))]>, - Requires<[HasV4T]>; + [(set IntRegs:$dst, (HexagonCONST32 tglobaladdr:$src1))]>; // Transfer a block address into a register def : Pat<(HexagonCONST32_GP tblockaddress:$src1), - (TFRI_V4 tblockaddress:$src1)>, - Requires<[HasV4T]>; - -let isExtended = 1, opExtendable = 2, AddedComplexity=50, -hasSideEffects = 0, isPredicated = 1, validSubTargets = HasV4SubT in -def TFRI_cPt_V4 : ALU32_ri<(outs IntRegs:$dst), - (ins PredRegs:$src1, s16Ext:$src2), - "if($src1) $dst = #$src2", - []>, - Requires<[HasV4T]>; - -let isExtended = 1, opExtendable = 2, AddedComplexity=50, isPredicatedFalse = 1, -hasSideEffects = 0, isPredicated = 1, validSubTargets = HasV4SubT in -def TFRI_cNotPt_V4 : ALU32_ri<(outs IntRegs:$dst), - (ins PredRegs:$src1, s16Ext:$src2), - "if(!$src1) $dst = #$src2", - []>, - Requires<[HasV4T]>; - -let isExtended = 1, opExtendable = 2, AddedComplexity=50, -hasSideEffects = 0, isPredicated = 1, validSubTargets = HasV4SubT in -def TFRI_cdnPt_V4 : ALU32_ri<(outs IntRegs:$dst), - (ins PredRegs:$src1, s16Ext:$src2), - "if($src1.new) $dst = #$src2", - []>, - Requires<[HasV4T]>; - -let isExtended = 1, opExtendable = 2, AddedComplexity=50, isPredicatedFalse = 1, -hasSideEffects = 0, isPredicated = 1, validSubTargets = HasV4SubT in -def TFRI_cdnNotPt_V4 : ALU32_ri<(outs IntRegs:$dst), - (ins PredRegs:$src1, s16Ext:$src2), - "if(!$src1.new) $dst = #$src2", - []>, - Requires<[HasV4T]>; - -let AddedComplexity = 50, Predicates = [HasV4T] in -def : Pat<(HexagonCONST32_GP tglobaladdr:$src1), - (TFRI_V4 tglobaladdr:$src1)>, - Requires<[HasV4T]>; - - -// Load - Indirect with long offset: These instructions take global address -// as an operand -let isExtended = 1, opExtendable = 3, AddedComplexity = 40, -validSubTargets = HasV4SubT in -def LDrid_ind_lo_V4 : LDInst<(outs DoubleRegs:$dst), - (ins IntRegs:$src1, u2Imm:$src2, globaladdressExt:$offset), - "$dst=memd($src1<<#$src2+##$offset)", - [(set (i64 DoubleRegs:$dst), - (load (add (shl IntRegs:$src1, u2ImmPred:$src2), - (HexagonCONST32 tglobaladdr:$offset))))]>, - Requires<[HasV4T]>; - -let AddedComplexity = 40 in -multiclass LD_indirect_lo { -let isExtended = 1, opExtendable = 3, validSubTargets = HasV4SubT in - def _lo_V4 : LDInst<(outs IntRegs:$dst), - (ins IntRegs:$src1, u2Imm:$src2, globaladdressExt:$offset), - !strconcat("$dst = ", - !strconcat(OpcStr, "($src1<<#$src2+##$offset)")), - [(set IntRegs:$dst, - (i32 (OpNode (add (shl IntRegs:$src1, u2ImmPred:$src2), - (HexagonCONST32 tglobaladdr:$offset)))))]>, - Requires<[HasV4T]>; -} - -defm LDrib_ind : LD_indirect_lo<"memb", sextloadi8>; -defm LDriub_ind : LD_indirect_lo<"memub", zextloadi8>; -defm LDriub_ind_anyext : LD_indirect_lo<"memub", extloadi8>; -defm LDrih_ind : LD_indirect_lo<"memh", sextloadi16>; -defm LDriuh_ind : LD_indirect_lo<"memuh", zextloadi16>; -defm LDriuh_ind_anyext : LD_indirect_lo<"memuh", extloadi16>; -defm LDriw_ind : LD_indirect_lo<"memw", load>; - -let AddedComplexity = 40 in -def : Pat <(i32 (sextloadi8 (add IntRegs:$src1, - (NumUsesBelowThresCONST32 tglobaladdr:$offset)))), - (i32 (LDrib_ind_lo_V4 IntRegs:$src1, 0, tglobaladdr:$offset))>, - Requires<[HasV4T]>; - -let AddedComplexity = 40 in -def : Pat <(i32 (zextloadi8 (add IntRegs:$src1, - (NumUsesBelowThresCONST32 tglobaladdr:$offset)))), - (i32 (LDriub_ind_lo_V4 IntRegs:$src1, 0, tglobaladdr:$offset))>, - Requires<[HasV4T]>; - -let Predicates = [HasV4T], AddedComplexity = 30 in { -def : Pat<(truncstorei8 (i32 IntRegs:$src1), u0AlwaysExtPred:$src2), - (S2_storerbabs u0AlwaysExtPred:$src2, IntRegs: $src1)>; - -def : Pat<(truncstorei16 (i32 IntRegs:$src1), u0AlwaysExtPred:$src2), - (S2_storerhabs u0AlwaysExtPred:$src2, IntRegs: $src1)>; - -def : Pat<(store (i32 IntRegs:$src1), u0AlwaysExtPred:$src2), - (S2_storeriabs u0AlwaysExtPred:$src2, IntRegs: $src1)>; -} - -let Predicates = [HasV4T], AddedComplexity = 30 in { -def : Pat<(i32 (load u0AlwaysExtPred:$src)), - (L4_loadri_abs u0AlwaysExtPred:$src)>; - -def : Pat<(i32 (sextloadi8 u0AlwaysExtPred:$src)), - (L4_loadrb_abs u0AlwaysExtPred:$src)>; - -def : Pat<(i32 (zextloadi8 u0AlwaysExtPred:$src)), - (L4_loadrub_abs u0AlwaysExtPred:$src)>; - -def : Pat<(i32 (sextloadi16 u0AlwaysExtPred:$src)), - (L4_loadrh_abs u0AlwaysExtPred:$src)>; - -def : Pat<(i32 (zextloadi16 u0AlwaysExtPred:$src)), - (L4_loadruh_abs u0AlwaysExtPred:$src)>; -} - -// Indexed store word - global address. -// memw(Rs+#u6:2)=#S8 -let AddedComplexity = 10 in -def STriw_offset_ext_V4 : STInst<(outs), - (ins IntRegs:$src1, u6_2Imm:$src2, globaladdress:$src3), - "memw($src1+#$src2) = ##$src3", - [(store (HexagonCONST32 tglobaladdr:$src3), - (add IntRegs:$src1, u6_2ImmPred:$src2))]>, - Requires<[HasV4T]>; - -def : Pat<(i64 (ctlz (i64 DoubleRegs:$src1))), - (i64 (A4_combineir (i32 0), (i32 (CTLZ64_rr DoubleRegs:$src1))))>, - Requires<[HasV4T]>; - -def : Pat<(i64 (cttz (i64 DoubleRegs:$src1))), - (i64 (A4_combineir (i32 0), (i32 (CTTZ64_rr DoubleRegs:$src1))))>, - Requires<[HasV4T]>; - - -// i8 -> i64 loads -// We need a complexity of 120 here to override preceding handling of -// zextloadi8. -let Predicates = [HasV4T], AddedComplexity = 120 in { -def: Pat <(i64 (extloadi8 (NumUsesBelowThresCONST32 tglobaladdr:$addr))), - (i64 (A4_combineir 0, (L4_loadrb_abs tglobaladdr:$addr)))>; + (TFRI_V4 tblockaddress:$src1)>; -def: Pat <(i64 (zextloadi8 (NumUsesBelowThresCONST32 tglobaladdr:$addr))), - (i64 (A4_combineir 0, (L4_loadrub_abs tglobaladdr:$addr)))>; - -def: Pat <(i64 (sextloadi8 (NumUsesBelowThresCONST32 tglobaladdr:$addr))), - (i64 (A2_sxtw (L4_loadrb_abs tglobaladdr:$addr)))>; - -def: Pat <(i64 (extloadi8 FoldGlobalAddr:$addr)), - (i64 (A4_combineir 0, (L4_loadrb_abs FoldGlobalAddr:$addr)))>; - -def: Pat <(i64 (zextloadi8 FoldGlobalAddr:$addr)), - (i64 (A4_combineir 0, (L4_loadrub_abs FoldGlobalAddr:$addr)))>; +let AddedComplexity = 50 in +def : Pat<(HexagonCONST32_GP tglobaladdr:$src1), + (TFRI_V4 tglobaladdr:$src1)>; -def: Pat <(i64 (sextloadi8 FoldGlobalAddr:$addr)), - (i64 (A2_sxtw (L4_loadrb_abs FoldGlobalAddr:$addr)))>; -} -// i16 -> i64 loads +// i8/i16/i32 -> i64 loads // We need a complexity of 120 here to override preceding handling of -// zextloadi16. +// zextload. let AddedComplexity = 120 in { -def: Pat <(i64 (extloadi16 (NumUsesBelowThresCONST32 tglobaladdr:$addr))), - (i64 (A4_combineir 0, (L4_loadrh_abs tglobaladdr:$addr)))>, - Requires<[HasV4T]>; + def: Loadam_pat; + def: Loadam_pat; + def: Loadam_pat; -def: Pat <(i64 (zextloadi16 (NumUsesBelowThresCONST32 tglobaladdr:$addr))), - (i64 (A4_combineir 0, (L4_loadruh_abs tglobaladdr:$addr)))>, - Requires<[HasV4T]>; + def: Loadam_pat; + def: Loadam_pat; + def: Loadam_pat; -def: Pat <(i64 (sextloadi16 (NumUsesBelowThresCONST32 tglobaladdr:$addr))), - (i64 (A2_sxtw (L4_loadrh_abs tglobaladdr:$addr)))>, - Requires<[HasV4T]>; - -def: Pat <(i64 (extloadi16 FoldGlobalAddr:$addr)), - (i64 (A4_combineir 0, (L4_loadrh_abs FoldGlobalAddr:$addr)))>, - Requires<[HasV4T]>; - -def: Pat <(i64 (zextloadi16 FoldGlobalAddr:$addr)), - (i64 (A4_combineir 0, (L4_loadruh_abs FoldGlobalAddr:$addr)))>, - Requires<[HasV4T]>; - -def: Pat <(i64 (sextloadi16 FoldGlobalAddr:$addr)), - (i64 (A2_sxtw (L4_loadrh_abs FoldGlobalAddr:$addr)))>, - Requires<[HasV4T]>; + def: Loadam_pat; + def: Loadam_pat; + def: Loadam_pat; } -// i32->i64 loads -// We need a complexity of 120 here to override preceding handling of -// zextloadi32. -let AddedComplexity = 120 in { -def: Pat <(i64 (extloadi32 (NumUsesBelowThresCONST32 tglobaladdr:$addr))), - (i64 (A4_combineir 0, (L4_loadri_abs tglobaladdr:$addr)))>, - Requires<[HasV4T]>; - -def: Pat <(i64 (zextloadi32 (NumUsesBelowThresCONST32 tglobaladdr:$addr))), - (i64 (A4_combineir 0, (L4_loadri_abs tglobaladdr:$addr)))>, - Requires<[HasV4T]>; -def: Pat <(i64 (sextloadi32 (NumUsesBelowThresCONST32 tglobaladdr:$addr))), - (i64 (A2_sxtw (L4_loadri_abs tglobaladdr:$addr)))>, - Requires<[HasV4T]>; +let AddedComplexity = 100 in { + def: Loada_pat; + def: Loada_pat; + def: Loada_pat; -def: Pat <(i64 (extloadi32 FoldGlobalAddr:$addr)), - (i64 (A4_combineir 0, (L4_loadri_abs FoldGlobalAddr:$addr)))>, - Requires<[HasV4T]>; + def: Loada_pat; + def: Loada_pat; + def: Loada_pat; -def: Pat <(i64 (zextloadi32 FoldGlobalAddr:$addr)), - (i64 (A4_combineir 0, (L4_loadri_abs FoldGlobalAddr:$addr)))>, - Requires<[HasV4T]>; - -def: Pat <(i64 (sextloadi32 FoldGlobalAddr:$addr)), - (i64 (A2_sxtw (L4_loadri_abs FoldGlobalAddr:$addr)))>, - Requires<[HasV4T]>; + def: Loada_pat; + def: Loada_pat; } -// Indexed store double word - global address. -// memw(Rs+#u6:2)=#S8 -let AddedComplexity = 10 in -def STrih_offset_ext_V4 : STInst<(outs), - (ins IntRegs:$src1, u6_1Imm:$src2, globaladdress:$src3), - "memh($src1+#$src2) = ##$src3", - [(truncstorei16 (HexagonCONST32 tglobaladdr:$src3), - (add IntRegs:$src1, u6_1ImmPred:$src2))]>, - Requires<[HasV4T]>; -// Map from store(globaladdress + x) -> memd(#foo + x) -let AddedComplexity = 100 in -def : Pat<(store (i64 DoubleRegs:$src1), - FoldGlobalAddrGP:$addr), - (S2_storerdabs FoldGlobalAddrGP:$addr, (i64 DoubleRegs:$src1))>, - Requires<[HasV4T]>; - -def : Pat<(atomic_store_64 FoldGlobalAddrGP:$addr, - (i64 DoubleRegs:$src1)), - (S2_storerdabs FoldGlobalAddrGP:$addr, (i64 DoubleRegs:$src1))>, - Requires<[HasV4T]>; - -// Map from store(globaladdress + x) -> memb(#foo + x) -let AddedComplexity = 100 in -def : Pat<(truncstorei8 (i32 IntRegs:$src1), FoldGlobalAddrGP:$addr), - (S2_storerbabs FoldGlobalAddrGP:$addr, (i32 IntRegs:$src1))>, - Requires<[HasV4T]>; - -def : Pat<(atomic_store_8 FoldGlobalAddrGP:$addr, (i32 IntRegs:$src1)), - (S2_storerbabs FoldGlobalAddrGP:$addr, (i32 IntRegs:$src1))>, - Requires<[HasV4T]>; - -// Map from store(globaladdress + x) -> memh(#foo + x) -let AddedComplexity = 100 in -def : Pat<(truncstorei16 (i32 IntRegs:$src1), FoldGlobalAddrGP:$addr), - (S2_storerhabs FoldGlobalAddrGP:$addr, (i32 IntRegs:$src1))>, - Requires<[HasV4T]>; - -def : Pat<(atomic_store_16 FoldGlobalAddrGP:$addr, (i32 IntRegs:$src1)), - (S2_storerhabs FoldGlobalAddrGP:$addr, (i32 IntRegs:$src1))>, - Requires<[HasV4T]>; - -// Map from store(globaladdress + x) -> memw(#foo + x) -let AddedComplexity = 100 in -def : Pat<(store (i32 IntRegs:$src1), FoldGlobalAddrGP:$addr), - (S2_storeriabs FoldGlobalAddrGP:$addr, (i32 IntRegs:$src1))>, - Requires<[HasV4T]>; - -def : Pat<(atomic_store_32 FoldGlobalAddrGP:$addr, (i32 IntRegs:$src1)), - (S2_storeriabs FoldGlobalAddrGP:$addr, (i32 IntRegs:$src1))>, - Requires<[HasV4T]>; - -// Map from load(globaladdress + x) -> memd(#foo + x) -let AddedComplexity = 100 in -def : Pat<(i64 (load FoldGlobalAddrGP:$addr)), - (i64 (L4_loadrd_abs FoldGlobalAddrGP:$addr))>, - Requires<[HasV4T]>; - -def : Pat<(atomic_load_64 FoldGlobalAddrGP:$addr), - (i64 (L4_loadrd_abs FoldGlobalAddrGP:$addr))>, - Requires<[HasV4T]>; - -// Map from load(globaladdress + x) -> memb(#foo + x) -let AddedComplexity = 100 in -def : Pat<(i32 (extloadi8 FoldGlobalAddrGP:$addr)), - (i32 (L4_loadrb_abs FoldGlobalAddrGP:$addr))>, - Requires<[HasV4T]>; - -// Map from load(globaladdress + x) -> memb(#foo + x) -let AddedComplexity = 100 in -def : Pat<(i32 (sextloadi8 FoldGlobalAddrGP:$addr)), - (i32 (L4_loadrb_abs FoldGlobalAddrGP:$addr))>, - Requires<[HasV4T]>; - -//let AddedComplexity = 100 in -let AddedComplexity = 100 in -def : Pat<(i32 (extloadi16 FoldGlobalAddrGP:$addr)), - (i32 (L4_loadrh_abs FoldGlobalAddrGP:$addr))>, - Requires<[HasV4T]>; - -// Map from load(globaladdress + x) -> memh(#foo + x) -let AddedComplexity = 100 in -def : Pat<(i32 (sextloadi16 FoldGlobalAddrGP:$addr)), - (i32 (L4_loadrh_abs FoldGlobalAddrGP:$addr))>, - Requires<[HasV4T]>; - -// Map from load(globaladdress + x) -> memuh(#foo + x) -let AddedComplexity = 100 in -def : Pat<(i32 (zextloadi16 FoldGlobalAddrGP:$addr)), - (i32 (L4_loadruh_abs FoldGlobalAddrGP:$addr))>, - Requires<[HasV4T]>; - -def : Pat<(atomic_load_16 FoldGlobalAddrGP:$addr), - (i32 (L4_loadruh_abs FoldGlobalAddrGP:$addr))>, - Requires<[HasV4T]>; - -// Map from load(globaladdress + x) -> memub(#foo + x) -let AddedComplexity = 100 in -def : Pat<(i32 (zextloadi8 FoldGlobalAddrGP:$addr)), - (i32 (L4_loadrub_abs FoldGlobalAddrGP:$addr))>, - Requires<[HasV4T]>; - -def : Pat<(atomic_load_8 FoldGlobalAddrGP:$addr), - (i32 (L4_loadrub_abs FoldGlobalAddrGP:$addr))>, - Requires<[HasV4T]>; - -// Map from load(globaladdress + x) -> memw(#foo + x) -let AddedComplexity = 100 in -def : Pat<(i32 (load FoldGlobalAddrGP:$addr)), - (i32 (L4_loadri_abs FoldGlobalAddrGP:$addr))>, - Requires<[HasV4T]>; +let AddedComplexity = 100 in { + def: Storea_pat; + def: Storea_pat; + def: Storea_pat; + def: Storea_pat; +} -def : Pat<(atomic_load_32 FoldGlobalAddrGP:$addr), - (i32 (L4_loadri_abs FoldGlobalAddrGP:$addr))>, - Requires<[HasV4T]>; +def: Loada_pat; +def: Loada_pat; +def: Loada_pat; +def: Loada_pat; + +def: Storea_pat, I32, addrgp, S2_storerbabs>; +def: Storea_pat, I32, addrgp, S2_storerhabs>; +def: Storea_pat, I32, addrgp, S2_storeriabs>; +def: Storea_pat, I64, addrgp, S2_storerdabs>; + +let Constraints = "@earlyclobber $dst" in +def Insert4 : PseudoM<(outs DoubleRegs:$dst), (ins IntRegs:$a, IntRegs:$b, + IntRegs:$c, IntRegs:$d), + ".error \"Should never try to emit Insert4\"", + [(set (i64 DoubleRegs:$dst), + (or (or (or (shl (i64 (zext (i32 (and (i32 IntRegs:$b), (i32 65535))))), + (i32 16)), + (i64 (zext (i32 (and (i32 IntRegs:$a), (i32 65535)))))), + (shl (i64 (anyext (i32 (and (i32 IntRegs:$c), (i32 65535))))), + (i32 32))), + (shl (i64 (anyext (i32 IntRegs:$d))), (i32 48))))]>; //===----------------------------------------------------------------------===// // :raw for of boundscheck:hi:lo insns //===----------------------------------------------------------------------===// // A4_boundscheck_lo: Detect if a register is within bounds. -let hasSideEffects = 0, isCodeGenOnly = 0 in +let hasSideEffects = 0 in def A4_boundscheck_lo: ALU64Inst < (outs PredRegs:$Pd), (ins DoubleRegs:$Rss, DoubleRegs:$Rtt), @@ -4146,7 +3945,7 @@ def A4_boundscheck_lo: ALU64Inst < } // A4_boundscheck_hi: Detect if a register is within bounds. -let hasSideEffects = 0, isCodeGenOnly = 0 in +let hasSideEffects = 0 in def A4_boundscheck_hi: ALU64Inst < (outs PredRegs:$Pd), (ins DoubleRegs:$Rss, DoubleRegs:$Rtt), @@ -4165,13 +3964,13 @@ def A4_boundscheck_hi: ALU64Inst < let Inst{12-8} = Rtt; } -let hasSideEffects = 0 in +let hasSideEffects = 0, isAsmParserOnly = 1 in def A4_boundscheck : MInst < (outs PredRegs:$Pd), (ins IntRegs:$Rs, DoubleRegs:$Rtt), "$Pd=boundscheck($Rs,$Rtt)">; // A4_tlbmatch: Detect if a VA/ASID matches a TLB entry. -let isPredicateLate = 1, hasSideEffects = 0, isCodeGenOnly = 0 in +let isPredicateLate = 1, hasSideEffects = 0 in def A4_tlbmatch : ALU64Inst<(outs PredRegs:$Pd), (ins DoubleRegs:$Rs, IntRegs:$Rt), "$Pd = tlbmatch($Rs, $Rt)", @@ -4198,7 +3997,7 @@ def HexagonDCFETCH : SDNode<"HexagonISD::DCFETCH", SDTHexagonDCFETCH, // Use LD0Inst for dcfetch, but set "mayLoad" to 0 because this doesn't // really do a load. -let hasSideEffects = 1, mayLoad = 0, isCodeGenOnly = 0 in +let hasSideEffects = 1, mayLoad = 0 in def Y2_dcfetchbo : LD0Inst<(outs), (ins IntRegs:$Rs, u11_3Imm:$u11_3), "dcfetch($Rs + #$u11_3)", [(HexagonDCFETCH IntRegs:$Rs, u11_3ImmPred:$u11_3)], @@ -4212,3 +4011,259 @@ def Y2_dcfetchbo : LD0Inst<(outs), (ins IntRegs:$Rs, u11_3Imm:$u11_3), let Inst{13} = 0b0; let Inst{10-0} = u11_3{13-3}; } + +//===----------------------------------------------------------------------===// +// Compound instructions +//===----------------------------------------------------------------------===// + +let isBranch = 1, hasSideEffects = 0, isExtentSigned = 1, + isPredicated = 1, isPredicatedNew = 1, isExtendable = 1, + opExtentBits = 11, opExtentAlign = 2, opExtendable = 1, + isTerminator = 1 in +class CJInst_tstbit_R0 + : InstHexagon<(outs), (ins IntRegs:$Rs, brtarget:$r9_2), + ""#px#" = tstbit($Rs, #0); if (" + #!if(np, "!","")#""#px#".new) jump:"#tnt#" $r9_2", + [], "", COMPOUND, TypeCOMPOUND>, OpcodeHexagon { + bits<4> Rs; + bits<11> r9_2; + + // np: !p[01] + let isPredicatedFalse = np; + // tnt: Taken/Not Taken + let isBrTaken = !if (!eq(tnt, "t"), "true", "false"); + let isTaken = !if (!eq(tnt, "t"), 1, 0); + + let IClass = 0b0001; + let Inst{27-26} = 0b00; + let Inst{25} = !if (!eq(px, "!p1"), 1, + !if (!eq(px, "p1"), 1, 0)); + let Inst{24-23} = 0b11; + let Inst{22} = np; + let Inst{21-20} = r9_2{10-9}; + let Inst{19-16} = Rs; + let Inst{13} = !if (!eq(tnt, "t"), 1, 0); + let Inst{9-8} = 0b11; + let Inst{7-1} = r9_2{8-2}; +} + +let Defs = [PC, P0], Uses = [P0] in { + def J4_tstbit0_tp0_jump_nt : CJInst_tstbit_R0<"p0", 0, "nt">; + def J4_tstbit0_tp0_jump_t : CJInst_tstbit_R0<"p0", 0, "t">; + def J4_tstbit0_fp0_jump_nt : CJInst_tstbit_R0<"p0", 1, "nt">; + def J4_tstbit0_fp0_jump_t : CJInst_tstbit_R0<"p0", 1, "t">; +} + +let Defs = [PC, P1], Uses = [P1] in { + def J4_tstbit0_tp1_jump_nt : CJInst_tstbit_R0<"p1", 0, "nt">; + def J4_tstbit0_tp1_jump_t : CJInst_tstbit_R0<"p1", 0, "t">; + def J4_tstbit0_fp1_jump_nt : CJInst_tstbit_R0<"p1", 1, "nt">; + def J4_tstbit0_fp1_jump_t : CJInst_tstbit_R0<"p1", 1, "t">; +} + + +let isBranch = 1, hasSideEffects = 0, + isExtentSigned = 1, isPredicated = 1, isPredicatedNew = 1, + isExtendable = 1, opExtentBits = 11, opExtentAlign = 2, + opExtendable = 2, isTerminator = 1 in +class CJInst_RR + : InstHexagon<(outs), (ins IntRegs:$Rs, IntRegs:$Rt, brtarget:$r9_2), + ""#px#" = cmp."#op#"($Rs, $Rt); if (" + #!if(np, "!","")#""#px#".new) jump:"#tnt#" $r9_2", + [], "", COMPOUND, TypeCOMPOUND>, OpcodeHexagon { + bits<4> Rs; + bits<4> Rt; + bits<11> r9_2; + + // np: !p[01] + let isPredicatedFalse = np; + // tnt: Taken/Not Taken + let isBrTaken = !if (!eq(tnt, "t"), "true", "false"); + let isTaken = !if (!eq(tnt, "t"), 1, 0); + + let IClass = 0b0001; + let Inst{27-23} = !if (!eq(op, "eq"), 0b01000, + !if (!eq(op, "gt"), 0b01001, + !if (!eq(op, "gtu"), 0b01010, 0))); + let Inst{22} = np; + let Inst{21-20} = r9_2{10-9}; + let Inst{19-16} = Rs; + let Inst{13} = !if (!eq(tnt, "t"), 1, 0); + // px: Predicate reg 0/1 + let Inst{12} = !if (!eq(px, "!p1"), 1, + !if (!eq(px, "p1"), 1, 0)); + let Inst{11-8} = Rt; + let Inst{7-1} = r9_2{8-2}; +} + +// P[10] taken/not taken. +multiclass T_tnt_CJInst_RR { + let Defs = [PC, P0], Uses = [P0] in { + def NAME#p0_jump_nt : CJInst_RR<"p0", op, np, "nt">; + def NAME#p0_jump_t : CJInst_RR<"p0", op, np, "t">; + } + let Defs = [PC, P1], Uses = [P1] in { + def NAME#p1_jump_nt : CJInst_RR<"p1", op, np, "nt">; + def NAME#p1_jump_t : CJInst_RR<"p1", op, np, "t">; + } +} +// Predicate / !Predicate +multiclass T_pnp_CJInst_RR{ + defm J4_cmp#NAME#_t : T_tnt_CJInst_RR; + defm J4_cmp#NAME#_f : T_tnt_CJInst_RR; +} +// TypeCJ Instructions compare RR and jump +defm eq : T_pnp_CJInst_RR<"eq">; +defm gt : T_pnp_CJInst_RR<"gt">; +defm gtu : T_pnp_CJInst_RR<"gtu">; + +let isBranch = 1, hasSideEffects = 0, isExtentSigned = 1, + isPredicated = 1, isPredicatedNew = 1, isExtendable = 1, opExtentBits = 11, + opExtentAlign = 2, opExtendable = 2, isTerminator = 1 in +class CJInst_RU5 + : InstHexagon<(outs), (ins IntRegs:$Rs, u5Imm:$U5, brtarget:$r9_2), + ""#px#" = cmp."#op#"($Rs, #$U5); if (" + #!if(np, "!","")#""#px#".new) jump:"#tnt#" $r9_2", + [], "", COMPOUND, TypeCOMPOUND>, OpcodeHexagon { + bits<4> Rs; + bits<5> U5; + bits<11> r9_2; + + // np: !p[01] + let isPredicatedFalse = np; + // tnt: Taken/Not Taken + let isBrTaken = !if (!eq(tnt, "t"), "true", "false"); + let isTaken = !if (!eq(tnt, "t"), 1, 0); + + let IClass = 0b0001; + let Inst{27-26} = 0b00; + // px: Predicate reg 0/1 + let Inst{25} = !if (!eq(px, "!p1"), 1, + !if (!eq(px, "p1"), 1, 0)); + let Inst{24-23} = !if (!eq(op, "eq"), 0b00, + !if (!eq(op, "gt"), 0b01, + !if (!eq(op, "gtu"), 0b10, 0))); + let Inst{22} = np; + let Inst{21-20} = r9_2{10-9}; + let Inst{19-16} = Rs; + let Inst{13} = !if (!eq(tnt, "t"), 1, 0); + let Inst{12-8} = U5; + let Inst{7-1} = r9_2{8-2}; +} +// P[10] taken/not taken. +multiclass T_tnt_CJInst_RU5 { + let Defs = [PC, P0], Uses = [P0] in { + def NAME#p0_jump_nt : CJInst_RU5<"p0", op, np, "nt">; + def NAME#p0_jump_t : CJInst_RU5<"p0", op, np, "t">; + } + let Defs = [PC, P1], Uses = [P1] in { + def NAME#p1_jump_nt : CJInst_RU5<"p1", op, np, "nt">; + def NAME#p1_jump_t : CJInst_RU5<"p1", op, np, "t">; + } +} +// Predicate / !Predicate +multiclass T_pnp_CJInst_RU5{ + defm J4_cmp#NAME#i_t : T_tnt_CJInst_RU5; + defm J4_cmp#NAME#i_f : T_tnt_CJInst_RU5; +} +// TypeCJ Instructions compare RI and jump +defm eq : T_pnp_CJInst_RU5<"eq">; +defm gt : T_pnp_CJInst_RU5<"gt">; +defm gtu : T_pnp_CJInst_RU5<"gtu">; + +let isBranch = 1, hasSideEffects = 0, isExtentSigned = 1, + isPredicated = 1, isPredicatedFalse = 1, isPredicatedNew = 1, + isExtendable = 1, opExtentBits = 11, opExtentAlign = 2, opExtendable = 1, + isTerminator = 1 in +class CJInst_Rn1 + : InstHexagon<(outs), (ins IntRegs:$Rs, brtarget:$r9_2), + ""#px#" = cmp."#op#"($Rs,#-1); if (" + #!if(np, "!","")#""#px#".new) jump:"#tnt#" $r9_2", + [], "", COMPOUND, TypeCOMPOUND>, OpcodeHexagon { + bits<4> Rs; + bits<11> r9_2; + + // np: !p[01] + let isPredicatedFalse = np; + // tnt: Taken/Not Taken + let isBrTaken = !if (!eq(tnt, "t"), "true", "false"); + let isTaken = !if (!eq(tnt, "t"), 1, 0); + + let IClass = 0b0001; + let Inst{27-26} = 0b00; + let Inst{25} = !if (!eq(px, "!p1"), 1, + !if (!eq(px, "p1"), 1, 0)); + + let Inst{24-23} = 0b11; + let Inst{22} = np; + let Inst{21-20} = r9_2{10-9}; + let Inst{19-16} = Rs; + let Inst{13} = !if (!eq(tnt, "t"), 1, 0); + let Inst{9-8} = !if (!eq(op, "eq"), 0b00, + !if (!eq(op, "gt"), 0b01, 0)); + let Inst{7-1} = r9_2{8-2}; +} + +// P[10] taken/not taken. +multiclass T_tnt_CJInst_Rn1 { + let Defs = [PC, P0], Uses = [P0] in { + def NAME#p0_jump_nt : CJInst_Rn1<"p0", op, np, "nt">; + def NAME#p0_jump_t : CJInst_Rn1<"p0", op, np, "t">; + } + let Defs = [PC, P1], Uses = [P1] in { + def NAME#p1_jump_nt : CJInst_Rn1<"p1", op, np, "nt">; + def NAME#p1_jump_t : CJInst_Rn1<"p1", op, np, "t">; + } +} +// Predicate / !Predicate +multiclass T_pnp_CJInst_Rn1{ + defm J4_cmp#NAME#n1_t : T_tnt_CJInst_Rn1; + defm J4_cmp#NAME#n1_f : T_tnt_CJInst_Rn1; +} +// TypeCJ Instructions compare -1 and jump +defm eq : T_pnp_CJInst_Rn1<"eq">; +defm gt : T_pnp_CJInst_Rn1<"gt">; + +// J4_jumpseti: Direct unconditional jump and set register to immediate. +let Defs = [PC], isBranch = 1, hasSideEffects = 0, hasNewValue = 1, + isExtentSigned = 1, opNewValue = 0, isExtendable = 1, opExtentBits = 11, + opExtentAlign = 2, opExtendable = 2 in +def J4_jumpseti: CJInst < + (outs IntRegs:$Rd), + (ins u6Imm:$U6, brtarget:$r9_2), + "$Rd = #$U6 ; jump $r9_2"> { + bits<4> Rd; + bits<6> U6; + bits<11> r9_2; + + let IClass = 0b0001; + let Inst{27-24} = 0b0110; + let Inst{21-20} = r9_2{10-9}; + let Inst{19-16} = Rd; + let Inst{13-8} = U6; + let Inst{7-1} = r9_2{8-2}; + } + +// J4_jumpsetr: Direct unconditional jump and transfer register. +let Defs = [PC], isBranch = 1, hasSideEffects = 0, hasNewValue = 1, + isExtentSigned = 1, opNewValue = 0, isExtendable = 1, opExtentBits = 11, + opExtentAlign = 2, opExtendable = 2 in +def J4_jumpsetr: CJInst < + (outs IntRegs:$Rd), + (ins IntRegs:$Rs, brtarget:$r9_2), + "$Rd = $Rs ; jump $r9_2"> { + bits<4> Rd; + bits<4> Rs; + bits<11> r9_2; + + let IClass = 0b0001; + let Inst{27-24} = 0b0111; + let Inst{21-20} = r9_2{10-9}; + let Inst{11-8} = Rd; + let Inst{19-16} = Rs; + let Inst{7-1} = r9_2{8-2}; + } + +// Duplex instructions +//===----------------------------------------------------------------------===// +include "HexagonIsetDx.td"