From 19c14abf1c4ccebfa7d07bdd6ea8462a15c0b749 Mon Sep 17 00:00:00 2001 From: Joey Gouly Date: Wed, 17 Jul 2013 13:59:38 +0000 Subject: [ARMv8] Add support for the NEON instructions vmaxnm/vminnm. This adds a new class for non-predicable NEON instructions and a new DecoderNamespace for v8 NEON instructions. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@186504 91177308-0d34-0410-b5e6-96231b3b80d8 --- include/llvm/IR/IntrinsicsARM.td | 2 ++ lib/Target/ARM/ARMCodeEmitter.cpp | 2 ++ lib/Target/ARM/ARMInstrFormats.td | 41 +++++++++++++++++++++ lib/Target/ARM/ARMInstrNEON.td | 45 ++++++++++++++++++++++++ lib/Target/ARM/Disassembler/ARMDisassembler.cpp | 18 ++++++++++ lib/Target/ARM/MCTargetDesc/ARMMCCodeEmitter.cpp | 13 +++++++ 6 files changed, 121 insertions(+) diff --git a/include/llvm/IR/IntrinsicsARM.td b/include/llvm/IR/IntrinsicsARM.td index 6cbd1c4..5ffea24 100644 --- a/include/llvm/IR/IntrinsicsARM.td +++ b/include/llvm/IR/IntrinsicsARM.td @@ -180,10 +180,12 @@ let Properties = [IntrNoMem, Commutative] in { // Vector Maximum. def int_arm_neon_vmaxs : Neon_2Arg_Intrinsic; def int_arm_neon_vmaxu : Neon_2Arg_Intrinsic; + def int_arm_neon_vmaxnm : Neon_2Arg_Intrinsic; // Vector Minimum. def int_arm_neon_vmins : Neon_2Arg_Intrinsic; def int_arm_neon_vminu : Neon_2Arg_Intrinsic; + def int_arm_neon_vminnm : Neon_2Arg_Intrinsic; // Vector Reciprocal Step. def int_arm_neon_vrecps : Neon_2Arg_Intrinsic; diff --git a/lib/Target/ARM/ARMCodeEmitter.cpp b/lib/Target/ARM/ARMCodeEmitter.cpp index e6f69a4..568ca85 100644 --- a/lib/Target/ARM/ARMCodeEmitter.cpp +++ b/lib/Target/ARM/ARMCodeEmitter.cpp @@ -167,6 +167,8 @@ namespace { const { return 0; } unsigned NEONThumb2DupPostEncoder(const MachineInstr &MI,unsigned Val) const { return 0; } + unsigned NEONThumb2V8PostEncoder(const MachineInstr &MI,unsigned Val) + const { return 0; } unsigned VFPThumb2PostEncoder(const MachineInstr&MI, unsigned Val) const { return 0; } unsigned getAdrLabelOpValue(const MachineInstr &MI, unsigned Op) diff --git a/lib/Target/ARM/ARMInstrFormats.td b/lib/Target/ARM/ARMInstrFormats.td index 9a542b9..a835187 100644 --- a/lib/Target/ARM/ARMInstrFormats.td +++ b/lib/Target/ARM/ARMInstrFormats.td @@ -1830,6 +1830,21 @@ class NeonXI pattern> + : InstARM { + let OutOperandList = oops; + let InOperandList = iops; + let AsmString = !strconcat(opc, ".", dt, "\t", asm); + let Pattern = pattern; + list Predicates = [HasNEON]; + let DecoderNamespace = "NEON"; + + let Inst{31-28} = 0b1111; +} + class NLdSt op21_20, bits<4> op11_8, bits<4> op7_4, dag oops, dag iops, InstrItinClass itin, string opc, string dt, string asm, string cstr, list pattern> @@ -2030,6 +2045,32 @@ class N3V op21_20, bits<4> op11_8, bit op6, bit op4, let Inst{5} = Vm{4}; } +class N3Vnp op27_23, bits<2> op21_20, bits<4> op11_8, bit op6, + bit op4, dag oops, dag iops,Format f, InstrItinClass itin, + string OpcodeStr, string Dt, ValueType ResTy, ValueType OpTy, + SDPatternOperator IntOp, bit Commutable, list pattern> + : NeonInp { + bits<5> Vd; + bits<5> Vn; + bits<5> Vm; + + // Encode instruction operands + let Inst{22} = Vd{4}; + let Inst{15-12} = Vd{3-0}; + let Inst{19-16} = Vn{3-0}; + let Inst{7} = Vn{4}; + let Inst{5} = Vm{4}; + let Inst{3-0} = Vm{3-0}; + + // Encode constant bits + let Inst{27-23} = op27_23; + let Inst{21-20} = op21_20; + let Inst{11-8} = op11_8; + let Inst{6} = op6; + let Inst{4} = op4; +} + class N3VLane32 op21_20, bits<4> op11_8, bit op6, bit op4, dag oops, dag iops, Format f, InstrItinClass itin, string opc, string dt, string asm, string cstr, diff --git a/lib/Target/ARM/ARMInstrNEON.td b/lib/Target/ARM/ARMInstrNEON.td index 32b8d67..f349f10 100644 --- a/lib/Target/ARM/ARMInstrNEON.td +++ b/lib/Target/ARM/ARMInstrNEON.td @@ -2541,6 +2541,16 @@ class N3VDInt op21_20, bits<4> op11_8, bit op4, let TwoOperandAliasConstraint = "$Vn = $Vd"; let isCommutable = Commutable; } + +class N3VDIntnp op27_23, bits<2> op21_20, bits<4> op11_8, bit op6, + bit op4, Format f, InstrItinClass itin, string OpcodeStr, + string Dt, ValueType ResTy, ValueType OpTy, + SDPatternOperator IntOp, bit Commutable> + : N3Vnp; + class N3VDIntSL op21_20, bits<4> op11_8, InstrItinClass itin, string OpcodeStr, string Dt, ValueType Ty, SDPatternOperator IntOp> : N3VLane32<0, 1, op21_20, op11_8, 1, 0, @@ -2552,6 +2562,7 @@ class N3VDIntSL op21_20, bits<4> op11_8, InstrItinClass itin, imm:$lane)))))]> { let isCommutable = 0; } + class N3VDIntSL16 op21_20, bits<4> op11_8, InstrItinClass itin, string OpcodeStr, string Dt, ValueType Ty, SDPatternOperator IntOp> : N3VLane16<0, 1, op21_20, op11_8, 1, 0, @@ -2584,6 +2595,16 @@ class N3VQInt op21_20, bits<4> op11_8, bit op4, let TwoOperandAliasConstraint = "$Vn = $Vd"; let isCommutable = Commutable; } + +class N3VQIntnp op27_23, bits<2> op21_20, bits<4> op11_8, bit op6, + bit op4, Format f, InstrItinClass itin, string OpcodeStr, + string Dt, ValueType ResTy, ValueType OpTy, + SDPatternOperator IntOp, bit Commutable> + : N3Vnp; + class N3VQIntSL op21_20, bits<4> op11_8, InstrItinClass itin, string OpcodeStr, string Dt, ValueType ResTy, ValueType OpTy, SDPatternOperator IntOp> @@ -4659,6 +4680,18 @@ def VMAXfq : N3VQInt<0, 0, 0b00, 0b1111, 0, N3RegFrm, IIC_VBINQ, "vmax", "f32", v4f32, v4f32, int_arm_neon_vmaxs, 1>; +// VMAXNM +let PostEncoderMethod = "NEONThumb2V8PostEncoder", DecoderNamespace = "v8NEON" in { + def VMAXNMND : N3VDIntnp<0b000110, 0b00, 0b1111, 0, 1, + N3RegFrm, NoItinerary, "vmaxnm", "f32", + v2f32, v2f32, int_arm_neon_vmaxnm, 1>, + Requires<[HasV8, HasNEON]>; + def VMAXNMNQ : N3VQIntnp<0b00110, 0b00, 0b1111, 1, 1, + N3RegFrm, NoItinerary, "vmaxnm", "f32", + v4f32, v4f32, int_arm_neon_vmaxnm, 1>, + Requires<[HasV8, HasNEON]>; +} + // VMIN : Vector Minimum defm VMINs : N3VInt_QHS<0, 0, 0b0110, 1, N3RegFrm, IIC_VSUBi4D, IIC_VSUBi4D, IIC_VSUBi4Q, IIC_VSUBi4Q, @@ -4673,6 +4706,18 @@ def VMINfq : N3VQInt<0, 0, 0b10, 0b1111, 0, N3RegFrm, IIC_VBINQ, "vmin", "f32", v4f32, v4f32, int_arm_neon_vmins, 1>; +// VMINNM +let PostEncoderMethod = "NEONThumb2V8PostEncoder", DecoderNamespace = "v8NEON" in { + def VMINNMND : N3VDIntnp<0b00110, 0b10, 0b1111, 0, 1, + N3RegFrm, NoItinerary, "vminnm", "f32", + v2f32, v2f32, int_arm_neon_vminnm, 1>, + Requires<[HasV8, HasNEON]>; + def VMINNMNQ : N3VQIntnp<0b00110, 0b10, 0b1111, 1, 1, + N3RegFrm, NoItinerary, "vminnm", "f32", + v4f32, v4f32, int_arm_neon_vminnm, 1>, + Requires<[HasV8, HasNEON]>; +} + // Vector Pairwise Operations. // VPADD : Vector Pairwise Add diff --git a/lib/Target/ARM/Disassembler/ARMDisassembler.cpp b/lib/Target/ARM/Disassembler/ARMDisassembler.cpp index 32ce3be..ce4dbb9 100644 --- a/lib/Target/ARM/Disassembler/ARMDisassembler.cpp +++ b/lib/Target/ARM/Disassembler/ARMDisassembler.cpp @@ -499,7 +499,14 @@ DecodeStatus ARMDisassembler::getInstruction(MCInst &MI, uint64_t &Size, } MI.clear(); + result = decodeInstruction(DecoderTablev8NEON32, MI, insn, Address, + this, STI); + if (result != MCDisassembler::Fail) { + Size = 4; + return result; + } + MI.clear(); Size = 0; return MCDisassembler::Fail; } @@ -818,6 +825,17 @@ DecodeStatus ThumbDisassembler::getInstruction(MCInst &MI, uint64_t &Size, } } + MI.clear(); + uint32_t NEONv8Insn = insn32; + NEONv8Insn &= 0xF3FFFFFF; // Clear bits 27-26 + result = decodeInstruction(DecoderTablev8NEON32, MI, NEONv8Insn, Address, + this, STI); + if (result != MCDisassembler::Fail) { + Size = 4; + return result; + } + + MI.clear(); Size = 0; return MCDisassembler::Fail; } diff --git a/lib/Target/ARM/MCTargetDesc/ARMMCCodeEmitter.cpp b/lib/Target/ARM/MCTargetDesc/ARMMCCodeEmitter.cpp index 8631d81..a18d465 100644 --- a/lib/Target/ARM/MCTargetDesc/ARMMCCodeEmitter.cpp +++ b/lib/Target/ARM/MCTargetDesc/ARMMCCodeEmitter.cpp @@ -315,6 +315,8 @@ public: unsigned EncodedValue) const; unsigned NEONThumb2DupPostEncoder(const MCInst &MI, unsigned EncodedValue) const; + unsigned NEONThumb2V8PostEncoder(const MCInst &MI, + unsigned EncodedValue) const; unsigned VFPThumb2PostEncoder(const MCInst &MI, unsigned EncodedValue) const; @@ -389,6 +391,17 @@ unsigned ARMMCCodeEmitter::NEONThumb2DupPostEncoder(const MCInst &MI, return EncodedValue; } +/// Post-process encoded NEON v8 instructions, and rewrite them to Thumb2 form +/// if we are in Thumb2. +unsigned ARMMCCodeEmitter::NEONThumb2V8PostEncoder(const MCInst &MI, + unsigned EncodedValue) const { + if (isThumb2()) { + EncodedValue |= 0xC000000; // Set bits 27-26 + } + + return EncodedValue; +} + /// VFPThumb2PostEncoder - Post-process encoded VFP instructions and rewrite /// them to their Thumb2 form if we are currently in Thumb2 mode. unsigned ARMMCCodeEmitter:: -- cgit v1.1