llvm/lib/Target/AMDGPU/VOPDInstructions.td - llvm-project - Git at Google

 //===-- VOPDInstructions.td - Vector Instruction Definitions --------------===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//

 //===----------------------------------------------------------------------===//
 // Encodings
 //===----------------------------------------------------------------------===//

 class VOPDe<bits<4> opX, bits<5> opY> : Enc64 {
   bits<9> src0X;
   bits<8> vsrc1X;
   bits<8> vdstX;
   bits<9> src0Y;
   bits<8> vsrc1Y;
   bits<8> vdstY;

   let Inst{8-0} = src0X;
   let Inst{16-9} = vsrc1X;
   let Inst{21-17} = opY;
   let Inst{25-22} = opX;
   let Inst{31-26} = 0x32; // encoding
   let Inst{40-32} = src0Y;
   let Inst{48-41} = vsrc1Y;
   let Inst{55-49} = vdstY{7-1};
   let Inst{63-56} = vdstX;
 }

 class VOPD_MADKe<bits<4> opX, bits<5> opY> : Enc96 {
   bits<9> src0X;
   bits<8> vsrc1X;
   bits<8> vdstX;
   bits<9> src0Y;
   bits<8> vsrc1Y;
   bits<8> vdstY;
   bits<32> imm;

   let Inst{8-0} = src0X;
   let Inst{16-9} = vsrc1X;
   let Inst{21-17} = opY;
   let Inst{25-22} = opX;
   let Inst{31-26} = 0x32; // encoding
   let Inst{40-32} = src0Y;
   let Inst{48-41} = vsrc1Y;
   let Inst{55-49} = vdstY{7-1};
   let Inst{63-56} = vdstX;
   let Inst{95-64} = imm;
 }

 //===----------------------------------------------------------------------===//
 // VOPD classes
 //===----------------------------------------------------------------------===//


 class GFXGenD<GFXGen Gen, list<string> DXPseudos, list<string> DYPseudos,
               Predicate subtargetPred = Gen.AssemblerPredicate> :
     GFXGen<Gen.AssemblerPredicate, Gen.DecoderNamespace, Gen.Suffix,
            Gen.Subtarget> {
   list<string> VOPDXPseudos = DXPseudos;
   list<string> VOPDYPseudos = DYPseudos;
   Predicate SubtargetPredicate = subtargetPred;
 }

 class VOPD_Base<dag outs, dag ins, string asm, VOP_Pseudo VDX, VOP_Pseudo VDY,
                 VOPD_Component XasVC, VOPD_Component YasVC, GFXGenD Gen>
     : VOPAnyCommon<outs, ins, asm, []>,
       VOP<NAME>,
       SIMCInstr<NAME, Gen.Subtarget> {
   // Fields for table indexing
   Instruction Opcode = !cast<Instruction>(NAME);
   bits<5> OpX = XasVC.VOPDOp;
   bits<5> OpY = YasVC.VOPDOp;
   bits<4> SubTgt = Gen.Subtarget;

   let VALU = 1;

   let DecoderNamespace = Gen.DecoderNamespace;
   let AssemblerPredicate = Gen.AssemblerPredicate;
   let WaveSizePredicate = isWave32;
   let isCodeGenOnly = 0;
   let SubtargetPredicate = Gen.SubtargetPredicate;
   let AsmMatchConverter  = "cvtVOPD";
   let Size = 8;
   let ReadsModeReg = !or(VDX.ReadsModeReg, VDY.ReadsModeReg);
   let mayRaiseFPException = ReadsModeReg;

   // V_DUAL_FMAC and V_DUAL_DOT2ACC_F32_F16 and V_DUAL_DOT2ACC_F32_BF16 need a
   // dummy src2 tied to dst for passes to track its uses. Its presence does not
   // affect VOPD formation rules because the rules for src2 and dst are the
   // same. src2X and src2Y should not be encoded.
   bit hasSrc2AccX = !or(!eq(VDX.Mnemonic, "v_fmac_f32"), !eq(VDX.Mnemonic, "v_dot2c_f32_f16"), !eq(VDX.Mnemonic, "v_dot2c_f32_bf16"));
   bit hasSrc2AccY = !or(!eq(VDY.Mnemonic, "v_fmac_f32"), !eq(VDY.Mnemonic, "v_dot2c_f32_f16"), !eq(VDY.Mnemonic, "v_dot2c_f32_bf16"));
   string ConstraintsX = !if(hasSrc2AccX, "$src2X = $vdstX", "");
   string ConstraintsY = !if(hasSrc2AccY, "$src2Y = $vdstY", "");
   let Constraints =
       ConstraintsX # !if(!and(hasSrc2AccX, hasSrc2AccY), ", ", "") # ConstraintsY;
   string DisableEncodingX = !if(hasSrc2AccX, "$src2X", "");
   string DisableEncodingY = !if(hasSrc2AccY, "$src2Y", "");
   let DisableEncoding =
       DisableEncodingX # !if(!and(hasSrc2AccX, hasSrc2AccY), ", ", "") # DisableEncodingY;

   let Uses = RegListUnion<VDX.Uses, VDY.Uses>.ret;
   let Defs = RegListUnion<VDX.Defs, VDY.Defs>.ret;
   let SchedRW = !listconcat(VDX.SchedRW, VDY.SchedRW);
 }

 class VOPD<dag outs, dag ins, string asm, VOP_Pseudo VDX, VOP_Pseudo VDY,
            VOPD_Component XasVC, VOPD_Component YasVC, GFXGenD Gen>
     : VOPD_Base<outs, ins, asm, VDX, VDY, XasVC, YasVC, Gen>,
       VOPDe<XasVC.VOPDOp{3-0}, YasVC.VOPDOp> {
   let Inst{16-9} = !if (!eq(VDX.Mnemonic, "v_mov_b32"), 0x0, vsrc1X);
   let Inst{48-41} = !if (!eq(VDY.Mnemonic, "v_mov_b32"), 0x0, vsrc1Y);
 }

 class VOPD_MADK<dag outs, dag ins, string asm, VOP_Pseudo VDX, VOP_Pseudo VDY,
                 VOPD_Component XasVC, VOPD_Component YasVC, GFXGenD Gen>
     : VOPD_Base<outs, ins, asm, VDX, VDY, XasVC, YasVC, Gen>,
       VOPD_MADKe<XasVC.VOPDOp{3-0}, YasVC.VOPDOp> {
   let Inst{16-9} = !if (!eq(VDX.Mnemonic, "v_mov_b32"), 0x0, vsrc1X);
   let Inst{48-41} = !if (!eq(VDY.Mnemonic, "v_mov_b32"), 0x0, vsrc1Y);
   let Size = 12;
   let FixedSize = 1;
 }

 defvar VOPDPseudosCommon = [
   "V_FMAC_F32_e32", "V_FMAAK_F32", "V_FMAMK_F32", "V_MUL_F32_e32",
   "V_ADD_F32_e32", "V_SUB_F32_e32", "V_SUBREV_F32_e32", "V_MUL_LEGACY_F32_e32",
   "V_MOV_B32_e32", "V_CNDMASK_B32_e32", "V_MAX_F32_e32", "V_MIN_F32_e32",
   "V_DOT2C_F32_F16_e32", "V_DOT2C_F32_BF16_e32"
 ];
 defvar VOPDYOnlyPseudosCommon = ["V_ADD_U32_e32", "V_LSHLREV_B32_e32",
                                  "V_AND_B32_e32"];

 defvar VOPDXPseudosGFX11 = VOPDPseudosCommon;
 defvar VOPDXPseudosGFX12 = VOPDPseudosCommon;
 defvar VOPDYPseudosGFX11 = !listconcat(VOPDXPseudosGFX11, VOPDYOnlyPseudosCommon);
 defvar VOPDYPseudosGFX12 = !listconcat(VOPDXPseudosGFX12, VOPDYOnlyPseudosCommon);

 def GFX11GenD : GFXGenD<GFX11Gen, VOPDXPseudosGFX11, VOPDYPseudosGFX11>;
 def GFX12GenD : GFXGenD<GFX12Gen, VOPDXPseudosGFX12, VOPDYPseudosGFX12>;


 def VOPDDstYOperand : RegisterOperand<VGPR_32, "printRegularOperand"> {
   let DecoderMethod = "decodeOperandVOPDDstY";
 }

 class getRenamed<string VOPDName, GFXGen Gen> {
   string ret = !if(!eq(Gen.Subtarget, GFX12Gen.Subtarget),
                    !if(!eq(VOPDName, "v_dual_max_f32"),
                        "v_dual_max_num_f32",
                        !if(!eq(VOPDName, "v_dual_min_f32"),
                            "v_dual_min_num_f32",
                            VOPDName)),
                    VOPDName);
 }

 foreach Gen = [GFX11GenD, GFX12GenD] in {
   foreach x = Gen.VOPDXPseudos in {
     foreach y = Gen.VOPDYPseudos in {
       defvar xInst = !cast<VOP_Pseudo>(x);
       defvar yInst = !cast<VOP_Pseudo>(y);
       defvar XasVC = !cast<VOPD_Component>(x);
       defvar YasVC = !cast<VOPD_Component>(y);
       defvar xAsmName = getRenamed<XasVC.VOPDName, Gen>.ret;
       defvar yAsmName = getRenamed<YasVC.VOPDName, Gen>.ret;
       defvar isMADK = !or(!eq(x, "V_FMAAK_F32"), !eq(x, "V_FMAMK_F32"),
                           !eq(y, "V_FMAAK_F32"), !eq(y, "V_FMAMK_F32"));
       // If X or Y is MADK (have a mandatory immediate), all src operands which
       // may contain an optional literal must use the VSrc_*_Deferred operand
       // type. Optional literal operands in MADK VOPD components always use this
       // operand form. If Both X and Y are MADK, the mandatory literal of X
       // additionally must use an alternate operand format which defers to the
       // 'real' Y literal
       defvar isOpXMADK = !or(!eq(x, "V_FMAAK_F32"), !eq(x, "V_FMAMK_F32"));
       defvar isOpYMADK = !or(!eq(y, "V_FMAAK_F32"), !eq(y, "V_FMAMK_F32"));
       defvar OpName = "V_DUAL_" # !substr(x,2) # "_X_" # !substr(y,2) # Gen.Suffix;
       defvar outs = (outs VGPRSrc_32:$vdstX, VOPDDstYOperand:$vdstY);
       if !or(isOpXMADK, isOpYMADK) then {
         if !and(isOpXMADK, isOpYMADK) then {
           defvar X_MADK_Pfl = !cast<VOP_MADK_Base>(xInst.Pfl);
           defvar ins = !con(xInst.Pfl.InsVOPDXDeferred, yInst.Pfl.InsVOPDY);
           defvar asm = xAsmName #" "# X_MADK_Pfl.AsmVOPDXDeferred #" :: "# yAsmName #" "# yInst.Pfl.AsmVOPDY;
           def OpName : VOPD_MADK<outs, ins, asm, xInst, yInst, XasVC, YasVC, Gen>;
         } else {
           defvar asm = xAsmName #" "# xInst.Pfl.AsmVOPDX #" :: "# yAsmName #" "# yInst.Pfl.AsmVOPDY;
           if isOpXMADK then {
             assert !not(isOpYMADK), "Expected only OpX as MADK";
             defvar ins = !con(xInst.Pfl.InsVOPDX, yInst.Pfl.InsVOPDYDeferred);
             def OpName : VOPD_MADK<outs, ins, asm, xInst, yInst, XasVC, YasVC, Gen>;
           } else {
             assert !not(isOpXMADK), "Expected only OpY as MADK";
             defvar ins = !con(xInst.Pfl.InsVOPDXDeferred, yInst.Pfl.InsVOPDY);
             def OpName : VOPD_MADK<outs, ins, asm, xInst, yInst, XasVC, YasVC, Gen>;
           }
         }
       } else {
         defvar ins = !con(xInst.Pfl.InsVOPDX, yInst.Pfl.InsVOPDY);
         defvar asm = xAsmName #" "# xInst.Pfl.AsmVOPDX #" :: "# yAsmName #" "# yInst.Pfl.AsmVOPDY;
         def OpName : VOPD<outs, ins, asm, xInst, yInst, XasVC, YasVC, Gen>;
       }
     }
   }
 }
	//===-- VOPDInstructions.td - Vector Instruction Definitions --------------===//
	//
	// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
	// See https://llvm.org/LICENSE.txt for license information.
	// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
	//
	//===----------------------------------------------------------------------===//

	//===----------------------------------------------------------------------===//
	// Encodings
	//===----------------------------------------------------------------------===//

	class VOPDe<bits<4> opX, bits<5> opY> : Enc64 {
	bits<9> src0X;
	bits<8> vsrc1X;
	bits<8> vdstX;
	bits<9> src0Y;
	bits<8> vsrc1Y;
	bits<8> vdstY;

	let Inst{8-0} = src0X;
	let Inst{16-9} = vsrc1X;
	let Inst{21-17} = opY;
	let Inst{25-22} = opX;
	let Inst{31-26} = 0x32; // encoding
	let Inst{40-32} = src0Y;
	let Inst{48-41} = vsrc1Y;
	let Inst{55-49} = vdstY{7-1};
	let Inst{63-56} = vdstX;
	}

	class VOPD_MADKe<bits<4> opX, bits<5> opY> : Enc96 {
	bits<9> src0X;
	bits<8> vsrc1X;
	bits<8> vdstX;
	bits<9> src0Y;
	bits<8> vsrc1Y;
	bits<8> vdstY;
	bits<32> imm;

	let Inst{8-0} = src0X;
	let Inst{16-9} = vsrc1X;
	let Inst{21-17} = opY;
	let Inst{25-22} = opX;
	let Inst{31-26} = 0x32; // encoding
	let Inst{40-32} = src0Y;
	let Inst{48-41} = vsrc1Y;
	let Inst{55-49} = vdstY{7-1};
	let Inst{63-56} = vdstX;
	let Inst{95-64} = imm;
	}

	//===----------------------------------------------------------------------===//
	// VOPD classes
	//===----------------------------------------------------------------------===//


	class GFXGenD<GFXGen Gen, list<string> DXPseudos, list<string> DYPseudos,
	Predicate subtargetPred = Gen.AssemblerPredicate> :
	GFXGen<Gen.AssemblerPredicate, Gen.DecoderNamespace, Gen.Suffix,
	Gen.Subtarget> {
	list<string> VOPDXPseudos = DXPseudos;
	list<string> VOPDYPseudos = DYPseudos;
	Predicate SubtargetPredicate = subtargetPred;
	}

	class VOPD_Base<dag outs, dag ins, string asm, VOP_Pseudo VDX, VOP_Pseudo VDY,
	VOPD_Component XasVC, VOPD_Component YasVC, GFXGenD Gen>
	: VOPAnyCommon<outs, ins, asm, []>,
	VOP<NAME>,
	SIMCInstr<NAME, Gen.Subtarget> {
	// Fields for table indexing
	Instruction Opcode = !cast<Instruction>(NAME);
	bits<5> OpX = XasVC.VOPDOp;
	bits<5> OpY = YasVC.VOPDOp;
	bits<4> SubTgt = Gen.Subtarget;

	let VALU = 1;

	let DecoderNamespace = Gen.DecoderNamespace;
	let AssemblerPredicate = Gen.AssemblerPredicate;
	let WaveSizePredicate = isWave32;
	let isCodeGenOnly = 0;
	let SubtargetPredicate = Gen.SubtargetPredicate;
	let AsmMatchConverter = "cvtVOPD";
	let Size = 8;
	let ReadsModeReg = !or(VDX.ReadsModeReg, VDY.ReadsModeReg);
	let mayRaiseFPException = ReadsModeReg;

	// V_DUAL_FMAC and V_DUAL_DOT2ACC_F32_F16 and V_DUAL_DOT2ACC_F32_BF16 need a
	// dummy src2 tied to dst for passes to track its uses. Its presence does not
	// affect VOPD formation rules because the rules for src2 and dst are the
	// same. src2X and src2Y should not be encoded.
	bit hasSrc2AccX = !or(!eq(VDX.Mnemonic, "v_fmac_f32"), !eq(VDX.Mnemonic, "v_dot2c_f32_f16"), !eq(VDX.Mnemonic, "v_dot2c_f32_bf16"));
	bit hasSrc2AccY = !or(!eq(VDY.Mnemonic, "v_fmac_f32"), !eq(VDY.Mnemonic, "v_dot2c_f32_f16"), !eq(VDY.Mnemonic, "v_dot2c_f32_bf16"));
	string ConstraintsX = !if(hasSrc2AccX, "$src2X = $vdstX", "");
	string ConstraintsY = !if(hasSrc2AccY, "$src2Y = $vdstY", "");
	let Constraints =
	ConstraintsX # !if(!and(hasSrc2AccX, hasSrc2AccY), ", ", "") # ConstraintsY;
	string DisableEncodingX = !if(hasSrc2AccX, "$src2X", "");
	string DisableEncodingY = !if(hasSrc2AccY, "$src2Y", "");
	let DisableEncoding =
	DisableEncodingX # !if(!and(hasSrc2AccX, hasSrc2AccY), ", ", "") # DisableEncodingY;

	let Uses = RegListUnion<VDX.Uses, VDY.Uses>.ret;
	let Defs = RegListUnion<VDX.Defs, VDY.Defs>.ret;
	let SchedRW = !listconcat(VDX.SchedRW, VDY.SchedRW);
	}

	class VOPD<dag outs, dag ins, string asm, VOP_Pseudo VDX, VOP_Pseudo VDY,
	VOPD_Component XasVC, VOPD_Component YasVC, GFXGenD Gen>
	: VOPD_Base<outs, ins, asm, VDX, VDY, XasVC, YasVC, Gen>,
	VOPDe<XasVC.VOPDOp{3-0}, YasVC.VOPDOp> {
	let Inst{16-9} = !if (!eq(VDX.Mnemonic, "v_mov_b32"), 0x0, vsrc1X);
	let Inst{48-41} = !if (!eq(VDY.Mnemonic, "v_mov_b32"), 0x0, vsrc1Y);
	}

	class VOPD_MADK<dag outs, dag ins, string asm, VOP_Pseudo VDX, VOP_Pseudo VDY,
	VOPD_Component XasVC, VOPD_Component YasVC, GFXGenD Gen>
	: VOPD_Base<outs, ins, asm, VDX, VDY, XasVC, YasVC, Gen>,
	VOPD_MADKe<XasVC.VOPDOp{3-0}, YasVC.VOPDOp> {
	let Inst{16-9} = !if (!eq(VDX.Mnemonic, "v_mov_b32"), 0x0, vsrc1X);
	let Inst{48-41} = !if (!eq(VDY.Mnemonic, "v_mov_b32"), 0x0, vsrc1Y);
	let Size = 12;
	let FixedSize = 1;
	}

	defvar VOPDPseudosCommon = [
	"V_FMAC_F32_e32", "V_FMAAK_F32", "V_FMAMK_F32", "V_MUL_F32_e32",
	"V_ADD_F32_e32", "V_SUB_F32_e32", "V_SUBREV_F32_e32", "V_MUL_LEGACY_F32_e32",
	"V_MOV_B32_e32", "V_CNDMASK_B32_e32", "V_MAX_F32_e32", "V_MIN_F32_e32",
	"V_DOT2C_F32_F16_e32", "V_DOT2C_F32_BF16_e32"
	];
	defvar VOPDYOnlyPseudosCommon = ["V_ADD_U32_e32", "V_LSHLREV_B32_e32",
	"V_AND_B32_e32"];

	defvar VOPDXPseudosGFX11 = VOPDPseudosCommon;
	defvar VOPDXPseudosGFX12 = VOPDPseudosCommon;
	defvar VOPDYPseudosGFX11 = !listconcat(VOPDXPseudosGFX11, VOPDYOnlyPseudosCommon);
	defvar VOPDYPseudosGFX12 = !listconcat(VOPDXPseudosGFX12, VOPDYOnlyPseudosCommon);

	def GFX11GenD : GFXGenD<GFX11Gen, VOPDXPseudosGFX11, VOPDYPseudosGFX11>;
	def GFX12GenD : GFXGenD<GFX12Gen, VOPDXPseudosGFX12, VOPDYPseudosGFX12>;


	def VOPDDstYOperand : RegisterOperand<VGPR_32, "printRegularOperand"> {
	let DecoderMethod = "decodeOperandVOPDDstY";
	}

	class getRenamed<string VOPDName, GFXGen Gen> {
	string ret = !if(!eq(Gen.Subtarget, GFX12Gen.Subtarget),
	!if(!eq(VOPDName, "v_dual_max_f32"),
	"v_dual_max_num_f32",
	!if(!eq(VOPDName, "v_dual_min_f32"),
	"v_dual_min_num_f32",
	VOPDName)),
	VOPDName);
	}

	foreach Gen = [GFX11GenD, GFX12GenD] in {
	foreach x = Gen.VOPDXPseudos in {
	foreach y = Gen.VOPDYPseudos in {
	defvar xInst = !cast<VOP_Pseudo>(x);
	defvar yInst = !cast<VOP_Pseudo>(y);
	defvar XasVC = !cast<VOPD_Component>(x);
	defvar YasVC = !cast<VOPD_Component>(y);
	defvar xAsmName = getRenamed<XasVC.VOPDName, Gen>.ret;
	defvar yAsmName = getRenamed<YasVC.VOPDName, Gen>.ret;
	defvar isMADK = !or(!eq(x, "V_FMAAK_F32"), !eq(x, "V_FMAMK_F32"),
	!eq(y, "V_FMAAK_F32"), !eq(y, "V_FMAMK_F32"));
	// If X or Y is MADK (have a mandatory immediate), all src operands which
	// may contain an optional literal must use the VSrc_*_Deferred operand
	// type. Optional literal operands in MADK VOPD components always use this
	// operand form. If Both X and Y are MADK, the mandatory literal of X
	// additionally must use an alternate operand format which defers to the
	// 'real' Y literal
	defvar isOpXMADK = !or(!eq(x, "V_FMAAK_F32"), !eq(x, "V_FMAMK_F32"));
	defvar isOpYMADK = !or(!eq(y, "V_FMAAK_F32"), !eq(y, "V_FMAMK_F32"));
	defvar OpName = "V_DUAL_" # !substr(x,2) # "_X_" # !substr(y,2) # Gen.Suffix;
	defvar outs = (outs VGPRSrc_32:$vdstX, VOPDDstYOperand:$vdstY);
	if !or(isOpXMADK, isOpYMADK) then {
	if !and(isOpXMADK, isOpYMADK) then {
	defvar X_MADK_Pfl = !cast<VOP_MADK_Base>(xInst.Pfl);
	defvar ins = !con(xInst.Pfl.InsVOPDXDeferred, yInst.Pfl.InsVOPDY);
	defvar asm = xAsmName #" "# X_MADK_Pfl.AsmVOPDXDeferred #" :: "# yAsmName #" "# yInst.Pfl.AsmVOPDY;
	def OpName : VOPD_MADK<outs, ins, asm, xInst, yInst, XasVC, YasVC, Gen>;
	} else {
	defvar asm = xAsmName #" "# xInst.Pfl.AsmVOPDX #" :: "# yAsmName #" "# yInst.Pfl.AsmVOPDY;
	if isOpXMADK then {
	assert !not(isOpYMADK), "Expected only OpX as MADK";
	defvar ins = !con(xInst.Pfl.InsVOPDX, yInst.Pfl.InsVOPDYDeferred);
	def OpName : VOPD_MADK<outs, ins, asm, xInst, yInst, XasVC, YasVC, Gen>;
	} else {
	assert !not(isOpXMADK), "Expected only OpY as MADK";
	defvar ins = !con(xInst.Pfl.InsVOPDXDeferred, yInst.Pfl.InsVOPDY);
	def OpName : VOPD_MADK<outs, ins, asm, xInst, yInst, XasVC, YasVC, Gen>;
	}
	}
	} else {
	defvar ins = !con(xInst.Pfl.InsVOPDX, yInst.Pfl.InsVOPDY);
	defvar asm = xAsmName #" "# xInst.Pfl.AsmVOPDX #" :: "# yAsmName #" "# yInst.Pfl.AsmVOPDY;
	def OpName : VOPD<outs, ins, asm, xInst, yInst, XasVC, YasVC, Gen>;
	}
	}
	}
	}