llvm/lib/Target/AMDGPU/VOPInstructions.td
Sam Kolton e88fc4046f [AMDGPU] SDWA: add support for GFX9 in peephole pass
Summary:
Added support based on merged SDWA pseudo instructions. Now peephole allow one scalar operand, omod and clamp modifiers.
Added several subtarget features for GFX9 SDWA.
This diff also contains changes from D34026.
Depends D34026

Reviewers: vpykhtin, rampitec, arsenm

Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye

Differential Revision: https://reviews.llvm.org/D34241

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@305986 91177308-0d34-0410-b5e6-96231b3b80d8
2017-06-22 06:26:41 +00:00

491 lines
15 KiB
TableGen

//===-- VOPInstructions.td - Vector Instruction Defintions ----------------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
// dummies for outer let
class LetDummies {
bit isCommutable;
bit isConvertibleToThreeAddress;
bit isMoveImm;
bit isReMaterializable;
bit isAsCheapAsAMove;
bit VOPAsmPrefer32Bit;
Predicate SubtargetPredicate;
string Constraints;
string DisableEncoding;
list<SchedReadWrite> SchedRW;
list<Register> Uses;
list<Register> Defs;
}
class VOP <string opName> {
string OpName = opName;
}
class VOPAnyCommon <dag outs, dag ins, string asm, list<dag> pattern> :
InstSI <outs, ins, asm, pattern> {
let mayLoad = 0;
let mayStore = 0;
let hasSideEffects = 0;
let UseNamedOperandTable = 1;
let VALU = 1;
let Uses = [EXEC];
}
class VOP3Common <dag outs, dag ins, string asm = "",
list<dag> pattern = [], bit HasMods = 0,
bit VOP3Only = 0> :
VOPAnyCommon <outs, ins, asm, pattern> {
// Using complex patterns gives VOP3 patterns a very high complexity rating,
// but standalone patterns are almost always preferred, so we need to adjust the
// priority lower. The goal is to use a high number to reduce complexity to
// zero (or less than zero).
let AddedComplexity = -1000;
let VOP3 = 1;
let AsmMatchConverter =
!if(!eq(VOP3Only,1),
"cvtVOP3",
!if(!eq(HasMods,1), "cvtVOP3_2_mod", ""));
let AsmVariantName = AMDGPUAsmVariants.VOP3;
let isCodeGenOnly = 0;
int Size = 8;
// Because SGPRs may be allowed if there are multiple operands, we
// need a post-isel hook to insert copies in order to avoid
// violating constant bus requirements.
let hasPostISelHook = 1;
}
class VOP3_Pseudo <string opName, VOPProfile P, list<dag> pattern = [],
bit VOP3Only = 0, bit isVOP3P = 0> :
InstSI <P.Outs64, !if(!and(isVOP3P, P.IsPacked), P.InsVOP3P, P.Ins64), "", pattern>,
VOP <opName>,
SIMCInstr<opName#"_e64", SIEncodingFamily.NONE>,
MnemonicAlias<opName#"_e64", opName> {
let isPseudo = 1;
let isCodeGenOnly = 1;
let UseNamedOperandTable = 1;
string Mnemonic = opName;
string AsmOperands = !if(!and(isVOP3P, P.IsPacked), P.AsmVOP3P, P.Asm64);
let Size = 8;
let mayLoad = 0;
let mayStore = 0;
let hasSideEffects = 0;
let SubtargetPredicate = isGCN;
// Because SGPRs may be allowed if there are multiple operands, we
// need a post-isel hook to insert copies in order to avoid
// violating constant bus requirements.
let hasPostISelHook = 1;
// Using complex patterns gives VOP3 patterns a very high complexity rating,
// but standalone patterns are almost always preferred, so we need to adjust the
// priority lower. The goal is to use a high number to reduce complexity to
// zero (or less than zero).
let AddedComplexity = -1000;
let VOP3 = 1;
let VALU = 1;
let FPClamp = P.HasFPClamp;
let Uses = [EXEC];
let AsmVariantName = AMDGPUAsmVariants.VOP3;
let AsmMatchConverter =
!if(!eq(VOP3Only,1),
!if(!and(P.IsPacked, isVOP3P), "cvtVOP3P", "cvtVOP3"),
!if(!eq(P.HasModifiers, 1),
"cvtVOP3_2_mod",
!if(!eq(P.HasOMod, 1), "cvtVOP3OMod", "")
)
);
VOPProfile Pfl = P;
}
class VOP3P_Pseudo <string opName, VOPProfile P, list<dag> pattern = []> :
VOP3_Pseudo<opName, P, pattern, 1, 1> {
let VOP3P = 1;
}
class VOP3_Real <VOP3_Pseudo ps, int EncodingFamily> :
InstSI <ps.OutOperandList, ps.InOperandList, ps.Mnemonic # ps.AsmOperands, []>,
SIMCInstr <ps.PseudoInstr, EncodingFamily> {
let isPseudo = 0;
let isCodeGenOnly = 0;
let UseNamedOperandTable = 1;
let Constraints = ps.Constraints;
let DisableEncoding = ps.DisableEncoding;
// copy relevant pseudo op flags
let SubtargetPredicate = ps.SubtargetPredicate;
let AsmMatchConverter = ps.AsmMatchConverter;
let AsmVariantName = ps.AsmVariantName;
let Constraints = ps.Constraints;
let DisableEncoding = ps.DisableEncoding;
let TSFlags = ps.TSFlags;
let UseNamedOperandTable = ps.UseNamedOperandTable;
let Uses = ps.Uses;
}
// XXX - Is there any reason to distingusih this from regular VOP3
// here?
class VOP3P_Real<VOP3P_Pseudo ps, int EncodingFamily> :
VOP3_Real<ps, EncodingFamily>;
class VOP3a<VOPProfile P> : Enc64 {
bits<2> src0_modifiers;
bits<9> src0;
bits<2> src1_modifiers;
bits<9> src1;
bits<2> src2_modifiers;
bits<9> src2;
bits<1> clamp;
bits<2> omod;
let Inst{8} = !if(P.HasSrc0Mods, src0_modifiers{1}, 0);
let Inst{9} = !if(P.HasSrc1Mods, src1_modifiers{1}, 0);
let Inst{10} = !if(P.HasSrc2Mods, src2_modifiers{1}, 0);
let Inst{31-26} = 0x34; //encoding
let Inst{40-32} = !if(P.HasSrc0, src0, 0);
let Inst{49-41} = !if(P.HasSrc1, src1, 0);
let Inst{58-50} = !if(P.HasSrc2, src2, 0);
let Inst{60-59} = !if(P.HasOMod, omod, 0);
let Inst{61} = !if(P.HasSrc0Mods, src0_modifiers{0}, 0);
let Inst{62} = !if(P.HasSrc1Mods, src1_modifiers{0}, 0);
let Inst{63} = !if(P.HasSrc2Mods, src2_modifiers{0}, 0);
}
class VOP3a_si <bits<9> op, VOPProfile P> : VOP3a<P> {
let Inst{25-17} = op;
let Inst{11} = !if(P.HasClamp, clamp{0}, 0);
}
class VOP3a_vi <bits<10> op, VOPProfile P> : VOP3a<P> {
let Inst{25-16} = op;
let Inst{15} = !if(P.HasClamp, clamp{0}, 0);
}
class VOP3e_si <bits<9> op, VOPProfile P> : VOP3a_si <op, P> {
bits<8> vdst;
let Inst{7-0} = !if(P.EmitDst, vdst{7-0}, 0);
}
class VOP3e_vi <bits<10> op, VOPProfile P> : VOP3a_vi <op, P> {
bits<8> vdst;
let Inst{7-0} = !if(P.EmitDst, vdst{7-0}, 0);
}
class VOP3be <VOPProfile P> : Enc64 {
bits<8> vdst;
bits<2> src0_modifiers;
bits<9> src0;
bits<2> src1_modifiers;
bits<9> src1;
bits<2> src2_modifiers;
bits<9> src2;
bits<7> sdst;
bits<2> omod;
let Inst{7-0} = vdst;
let Inst{14-8} = sdst;
let Inst{31-26} = 0x34; //encoding
let Inst{40-32} = !if(P.HasSrc0, src0, 0);
let Inst{49-41} = !if(P.HasSrc1, src1, 0);
let Inst{58-50} = !if(P.HasSrc2, src2, 0);
let Inst{60-59} = !if(P.HasOMod, omod, 0);
let Inst{61} = !if(P.HasSrc0Mods, src0_modifiers{0}, 0);
let Inst{62} = !if(P.HasSrc1Mods, src1_modifiers{0}, 0);
let Inst{63} = !if(P.HasSrc2Mods, src2_modifiers{0}, 0);
}
class VOP3Pe <bits<10> op, VOPProfile P> : Enc64 {
bits<8> vdst;
// neg, neg_hi, op_sel put in srcN_modifiers
bits<4> src0_modifiers;
bits<9> src0;
bits<4> src1_modifiers;
bits<9> src1;
bits<4> src2_modifiers;
bits<9> src2;
bits<1> clamp;
let Inst{7-0} = vdst;
let Inst{8} = !if(P.HasSrc0Mods, src0_modifiers{1}, 0); // neg_hi src0
let Inst{9} = !if(P.HasSrc1Mods, src1_modifiers{1}, 0); // neg_hi src1
let Inst{10} = !if(P.HasSrc2Mods, src2_modifiers{1}, 0); // neg_hi src2
let Inst{11} = !if(!and(P.HasSrc0, P.HasOpSel), src0_modifiers{2}, 0); // op_sel(0)
let Inst{12} = !if(!and(P.HasSrc1, P.HasOpSel), src1_modifiers{2}, 0); // op_sel(1)
let Inst{13} = !if(!and(P.HasSrc2, P.HasOpSel), src2_modifiers{2}, 0); // op_sel(2)
let Inst{14} = !if(!and(P.HasSrc2, P.HasOpSel), src2_modifiers{3}, 0); // op_sel_hi(2)
let Inst{15} = !if(P.HasClamp, clamp{0}, 0);
let Inst{25-16} = op;
let Inst{31-26} = 0x34; //encoding
let Inst{40-32} = !if(P.HasSrc0, src0, 0);
let Inst{49-41} = !if(P.HasSrc1, src1, 0);
let Inst{58-50} = !if(P.HasSrc2, src2, 0);
let Inst{59} = !if(!and(P.HasSrc0, P.HasOpSel), src0_modifiers{3}, 0); // op_sel_hi(0)
let Inst{60} = !if(!and(P.HasSrc1, P.HasOpSel), src1_modifiers{3}, 0); // op_sel_hi(1)
let Inst{61} = !if(P.HasSrc0Mods, src0_modifiers{0}, 0); // neg (lo)
let Inst{62} = !if(P.HasSrc1Mods, src1_modifiers{0}, 0); // neg (lo)
let Inst{63} = !if(P.HasSrc2Mods, src2_modifiers{0}, 0); // neg (lo)
}
class VOP3be_si <bits<9> op, VOPProfile P> : VOP3be<P> {
let Inst{25-17} = op;
}
class VOP3be_vi <bits<10> op, VOPProfile P> : VOP3be<P> {
bits<1> clamp;
let Inst{25-16} = op;
let Inst{15} = !if(P.HasClamp, clamp{0}, 0);
}
def SDWA {
// sdwa_sel
int BYTE_0 = 0;
int BYTE_1 = 1;
int BYTE_2 = 2;
int BYTE_3 = 3;
int WORD_0 = 4;
int WORD_1 = 5;
int DWORD = 6;
// dst_unused
int UNUSED_PAD = 0;
int UNUSED_SEXT = 1;
int UNUSED_PRESERVE = 2;
}
class VOP_SDWAe<VOPProfile P> : Enc64 {
bits<8> src0;
bits<3> src0_sel;
bits<2> src0_modifiers; // float: {abs,neg}, int {sext}
bits<3> src1_sel;
bits<2> src1_modifiers;
bits<3> dst_sel;
bits<2> dst_unused;
bits<1> clamp;
let Inst{39-32} = !if(P.HasSrc0, src0{7-0}, 0);
let Inst{42-40} = !if(P.EmitDst, dst_sel{2-0}, SDWA.DWORD);
let Inst{44-43} = !if(P.EmitDst, dst_unused{1-0}, SDWA.UNUSED_PRESERVE);
let Inst{45} = !if(P.HasSDWAClamp, clamp{0}, 0);
let Inst{50-48} = !if(P.HasSrc0, src0_sel{2-0}, SDWA.DWORD);
let Inst{51} = !if(P.HasSrc0IntMods, src0_modifiers{0}, 0);
let Inst{53-52} = !if(P.HasSrc0FloatMods, src0_modifiers{1-0}, 0);
let Inst{58-56} = !if(P.HasSrc1, src1_sel{2-0}, SDWA.DWORD);
let Inst{59} = !if(P.HasSrc1IntMods, src1_modifiers{0}, 0);
let Inst{61-60} = !if(P.HasSrc1FloatMods, src1_modifiers{1-0}, 0);
}
// GFX9 adds two features to SDWA:
// 1. Add 3 fields to the SDWA microcode word: S0, S1 and OMOD.
// a. S0 and S1 indicate that source 0 and 1 respectively are SGPRs rather
// than VGPRs (at most 1 can be an SGPR);
// b. OMOD is the standard output modifier (result *2, *4, /2)
// 2. Add a new version of the SDWA microcode word for VOPC: SDWAB. This
// replaces OMOD and the dest fields with SD and SDST (SGPR destination)
// field.
// a. When SD=1, the SDST is used as the destination for the compare result;
// b. When SD=0, VCC is used.
//
// In GFX9, V_MAC_F16, V_MAC_F32 opcodes cannot be used with SDWA
// gfx9 SDWA basic encoding
class VOP_SDWA9e<VOPProfile P> : Enc64 {
bits<9> src0; // {src0_sgpr{0}, src0{7-0}}
bits<3> src0_sel;
bits<2> src0_modifiers; // float: {abs,neg}, int {sext}
bits<3> src1_sel;
bits<2> src1_modifiers;
bits<1> src1_sgpr;
let Inst{39-32} = !if(P.HasSrc0, src0{7-0}, 0);
let Inst{50-48} = !if(P.HasSrc0, src0_sel{2-0}, SDWA.DWORD);
let Inst{51} = !if(P.HasSrc0IntMods, src0_modifiers{0}, 0);
let Inst{53-52} = !if(P.HasSrc0FloatMods, src0_modifiers{1-0}, 0);
let Inst{55} = !if(P.HasSrc0, src0{8}, 0);
let Inst{58-56} = !if(P.HasSrc1, src1_sel{2-0}, SDWA.DWORD);
let Inst{59} = !if(P.HasSrc1IntMods, src1_modifiers{0}, 0);
let Inst{61-60} = !if(P.HasSrc1FloatMods, src1_modifiers{1-0}, 0);
let Inst{63} = 0; // src1_sgpr - should be specified in subclass
}
// gfx9 SDWA-A
class VOP_SDWA9Ae<VOPProfile P> : VOP_SDWA9e<P> {
bits<3> dst_sel;
bits<2> dst_unused;
bits<1> clamp;
bits<2> omod;
let Inst{42-40} = !if(P.EmitDst, dst_sel{2-0}, SDWA.DWORD);
let Inst{44-43} = !if(P.EmitDst, dst_unused{1-0}, SDWA.UNUSED_PRESERVE);
let Inst{45} = !if(P.HasSDWAClamp, clamp{0}, 0);
let Inst{47-46} = !if(P.HasSDWAOMod, omod{1-0}, 0);
}
// gfx9 SDWA-B
class VOP_SDWA9Be<VOPProfile P> : VOP_SDWA9e<P> {
bits<8> sdst; // {vcc_sdst{0}, sdst{6-0}}
let Inst{46-40} = !if(P.EmitDst, sdst{6-0}, 0);
let Inst{47} = !if(P.EmitDst, sdst{7}, 0);
}
class VOP_SDWA_Pseudo <string opName, VOPProfile P, list<dag> pattern=[]> :
InstSI <P.OutsSDWA, P.InsSDWA, "", pattern>,
VOP <opName>,
SIMCInstr <opName#"_sdwa", SIEncodingFamily.NONE>,
MnemonicAlias <opName#"_sdwa", opName> {
let isPseudo = 1;
let isCodeGenOnly = 1;
let UseNamedOperandTable = 1;
string Mnemonic = opName;
string AsmOperands = P.AsmSDWA;
string AsmOperands9 = P.AsmSDWA9;
let Size = 8;
let mayLoad = 0;
let mayStore = 0;
let hasSideEffects = 0;
let VALU = 1;
let SDWA = 1;
let Uses = [EXEC];
let SubtargetPredicate = !if(P.HasExt, HasSDWA, DisableInst);
let AssemblerPredicate = !if(P.HasExt, HasSDWA, DisableInst);
let AsmVariantName = !if(P.HasExt, AMDGPUAsmVariants.SDWA,
AMDGPUAsmVariants.Disable);
let DecoderNamespace = "SDWA";
VOPProfile Pfl = P;
}
class VOP_SDWA_Real <VOP_SDWA_Pseudo ps> :
InstSI <ps.OutOperandList, ps.InOperandList, ps.Mnemonic # ps.AsmOperands, []>,
SIMCInstr <ps.PseudoInstr, SIEncodingFamily.SDWA> {
let isPseudo = 0;
let isCodeGenOnly = 0;
let Defs = ps.Defs;
let Uses = ps.Uses;
let SchedRW = ps.SchedRW;
let hasSideEffects = ps.hasSideEffects;
let Constraints = ps.Constraints;
let DisableEncoding = ps.DisableEncoding;
// Copy relevant pseudo op flags
let SubtargetPredicate = ps.SubtargetPredicate;
let AssemblerPredicate = ps.AssemblerPredicate;
let AsmMatchConverter = ps.AsmMatchConverter;
let AsmVariantName = ps.AsmVariantName;
let UseNamedOperandTable = ps.UseNamedOperandTable;
let DecoderNamespace = ps.DecoderNamespace;
let Constraints = ps.Constraints;
let DisableEncoding = ps.DisableEncoding;
let TSFlags = ps.TSFlags;
}
class VOP_SDWA9_Real <VOP_SDWA_Pseudo ps> :
InstSI <ps.OutOperandList, ps.InOperandList, ps.Mnemonic # ps.AsmOperands9, []>,
SIMCInstr <ps.PseudoInstr, SIEncodingFamily.SDWA9> {
let isPseudo = 0;
let isCodeGenOnly = 0;
let Defs = ps.Defs;
let Uses = ps.Uses;
let SchedRW = ps.SchedRW;
let hasSideEffects = ps.hasSideEffects;
let Constraints = ps.Constraints;
let DisableEncoding = ps.DisableEncoding;
let SubtargetPredicate = !if(ps.Pfl.HasSDWA9, HasSDWA9, DisableInst);
let AssemblerPredicate = !if(ps.Pfl.HasSDWA9, HasSDWA9, DisableInst);
let AsmVariantName = !if(ps.Pfl.HasSDWA9, AMDGPUAsmVariants.SDWA9,
AMDGPUAsmVariants.Disable);
let DecoderNamespace = "SDWA9";
// Copy relevant pseudo op flags
let AsmMatchConverter = ps.AsmMatchConverter;
let UseNamedOperandTable = ps.UseNamedOperandTable;
let Constraints = ps.Constraints;
let DisableEncoding = ps.DisableEncoding;
let TSFlags = ps.TSFlags;
}
class VOP_DPPe<VOPProfile P> : Enc64 {
bits<2> src0_modifiers;
bits<8> src0;
bits<2> src1_modifiers;
bits<9> dpp_ctrl;
bits<1> bound_ctrl;
bits<4> bank_mask;
bits<4> row_mask;
let Inst{39-32} = !if(P.HasSrc0, src0{7-0}, 0);
let Inst{48-40} = dpp_ctrl;
let Inst{51} = bound_ctrl;
let Inst{52} = !if(P.HasSrc0Mods, src0_modifiers{0}, 0); // src0_neg
let Inst{53} = !if(P.HasSrc0Mods, src0_modifiers{1}, 0); // src0_abs
let Inst{54} = !if(P.HasSrc1Mods, src1_modifiers{0}, 0); // src1_neg
let Inst{55} = !if(P.HasSrc1Mods, src1_modifiers{1}, 0); // src1_abs
let Inst{59-56} = bank_mask;
let Inst{63-60} = row_mask;
}
class VOP_DPP <string OpName, VOPProfile P> :
InstSI <P.OutsDPP, P.InsDPP, OpName#P.AsmDPP, []>,
VOP_DPPe<P> {
let mayLoad = 0;
let mayStore = 0;
let hasSideEffects = 0;
let UseNamedOperandTable = 1;
let VALU = 1;
let DPP = 1;
let Size = 8;
let AsmMatchConverter = !if(!eq(P.HasModifiers,1), "cvtDPP", "");
let SubtargetPredicate = HasDPP;
let AssemblerPredicate = !if(P.HasExt, HasDPP, DisableInst);
let AsmVariantName = !if(P.HasExt, AMDGPUAsmVariants.DPP,
AMDGPUAsmVariants.Disable);
let DecoderNamespace = "DPP";
}
include "VOPCInstructions.td"
include "VOP1Instructions.td"
include "VOP2Instructions.td"
include "VOP3Instructions.td"
include "VOP3PInstructions.td"