Files
clang-p2996/llvm/lib/Target/AMDGPU/SIInstrFormats.td
Ruiling Song 9119d9bfce AMDGPU/SIInsertWait: Skip dummy tied source
For D16 memory load instructions, the hardware usually only write to half
of the 32bit register, but we define the destination register using
32bit register for the MachineIR instruction. Without the extra tied
source register, LLVM framework will think previous write to the other
half of the register being dead. This is because by using 32bit register
as the destination register, LLVM will think the instruction will always
overwrite the whole 32bit register. By adding the extra tied source,
LLVM will think we are reading the register, so previous write to the
register will not be dead. This dummy tied source is introducing
unnecessary read-after-write dependency. The change here is to bypass the
tied source that can be skipped, thus avoiding an unnecessary s_waitcnt.

Reviewed by: foad

Differential Revision: https://reviews.llvm.org/D140537
2023-01-11 09:59:35 +08:00

470 lines
11 KiB
TableGen

//===-- SIInstrFormats.td - SI Instruction Encodings ----------------------===//
//
// 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
//
//===----------------------------------------------------------------------===//
//
// SI Instruction format definitions.
//
//===----------------------------------------------------------------------===//
class InstSI <dag outs, dag ins, string asm = "",
list<dag> pattern = []> :
AMDGPUInst<outs, ins, asm, pattern>, GCNPredicateControl {
// Low bits - basic encoding information.
field bit SALU = 0;
field bit VALU = 0;
// SALU instruction formats.
field bit SOP1 = 0;
field bit SOP2 = 0;
field bit SOPC = 0;
field bit SOPK = 0;
field bit SOPP = 0;
// VALU instruction formats.
field bit VOP1 = 0;
field bit VOP2 = 0;
field bit VOPC = 0;
field bit VOP3 = 0;
field bit VOP3P = 0;
field bit VINTRP = 0;
field bit SDWA = 0;
field bit DPP = 0;
field bit TRANS = 0;
// Memory instruction formats.
field bit MUBUF = 0;
field bit MTBUF = 0;
field bit SMRD = 0;
field bit MIMG = 0;
field bit EXP = 0;
field bit FLAT = 0;
field bit DS = 0;
// Pseudo instruction formats.
field bit VGPRSpill = 0;
field bit SGPRSpill = 0;
// LDSDIR instruction format.
field bit LDSDIR = 0;
// VINTERP instruction format.
field bit VINTERP = 0;
// High bits - other information.
field bit VM_CNT = 0;
field bit EXP_CNT = 0;
field bit LGKM_CNT = 0;
// Whether WQM _must_ be enabled for this instruction.
field bit WQM = 0;
// Whether WQM _must_ be disabled for this instruction.
field bit DisableWQM = 0;
field bit Gather4 = 0;
// Most sopk treat the immediate as a signed 16-bit, however some
// use it as unsigned.
field bit SOPKZext = 0;
// This is an s_store_dword* instruction that requires a cache flush
// on wave termination. It is necessary to distinguish from mayStore
// SMEM instructions like the cache flush ones.
field bit ScalarStore = 0;
// Whether the operands can be ignored when computing the
// instruction size.
field bit FixedSize = 0;
// This bit tells the assembler to use the 32-bit encoding in case it
// is unable to infer the encoding from the operands.
field bit VOPAsmPrefer32Bit = 0;
// This bit indicates that this is a VOP3 opcode which supports op_sel
// modifier.
field bit VOP3_OPSEL = 0;
// Is it possible for this instruction to be atomic?
field bit maybeAtomic = 0;
// This bit indicates that this is a VI instruction which is renamed
// in GFX9. Required for correct mapping from pseudo to MC.
field bit renamedInGFX9 = 0;
// This bit indicates that this has a floating point result type, so
// the clamp modifier has floating point semantics.
field bit FPClamp = 0;
// This bit indicates that instruction may support integer clamping
// which depends on GPU features.
field bit IntClamp = 0;
// This field indicates that the clamp applies to the low component
// of a packed output register.
field bit ClampLo = 0;
// This field indicates that the clamp applies to the high component
// of a packed output register.
field bit ClampHi = 0;
// This bit indicates that this is a packed VOP3P instruction
field bit IsPacked = 0;
// This bit indicates that this is a D16 buffer instruction.
field bit D16Buf = 0;
// This field indicates that FLAT instruction accesses FLAT_GLBL segment.
// Must be 0 for non-FLAT instructions.
field bit FlatGlobal = 0;
// Reads the mode register, usually for FP environment.
field bit ReadsModeReg = 0;
// This bit indicates that this uses the floating point double precision
// rounding mode flags
field bit FPDPRounding = 0;
// Instruction is FP atomic.
field bit FPAtomic = 0;
// This bit indicates that this is one of MFMA instructions.
field bit IsMAI = 0;
// This bit indicates that this is one of DOT instructions.
field bit IsDOT = 0;
// This field indicates that FLAT instruction accesses FLAT_SCRATCH segment.
// Must be 0 for non-FLAT instructions.
field bit FlatScratch = 0;
// Atomic without a return.
field bit IsAtomicNoRet = 0;
// Atomic with return.
field bit IsAtomicRet = 0;
// This bit indicates that this is one of WMMA instructions.
field bit IsWMMA = 0;
// This bit indicates that tied source will not be read.
field bit TiedSourceNotRead = 0;
// These need to be kept in sync with the enum in SIInstrFlags.
let TSFlags{0} = SALU;
let TSFlags{1} = VALU;
let TSFlags{2} = SOP1;
let TSFlags{3} = SOP2;
let TSFlags{4} = SOPC;
let TSFlags{5} = SOPK;
let TSFlags{6} = SOPP;
let TSFlags{7} = VOP1;
let TSFlags{8} = VOP2;
let TSFlags{9} = VOPC;
let TSFlags{10} = VOP3;
let TSFlags{12} = VOP3P;
let TSFlags{13} = VINTRP;
let TSFlags{14} = SDWA;
let TSFlags{15} = DPP;
let TSFlags{16} = TRANS;
let TSFlags{17} = MUBUF;
let TSFlags{18} = MTBUF;
let TSFlags{19} = SMRD;
let TSFlags{20} = MIMG;
let TSFlags{21} = EXP;
let TSFlags{22} = FLAT;
let TSFlags{23} = DS;
let TSFlags{24} = VGPRSpill;
let TSFlags{25} = SGPRSpill;
let TSFlags{26} = LDSDIR;
let TSFlags{27} = VINTERP;
let TSFlags{32} = VM_CNT;
let TSFlags{33} = EXP_CNT;
let TSFlags{34} = LGKM_CNT;
let TSFlags{35} = WQM;
let TSFlags{36} = DisableWQM;
let TSFlags{37} = Gather4;
let TSFlags{38} = SOPKZext;
let TSFlags{39} = ScalarStore;
let TSFlags{40} = FixedSize;
let TSFlags{41} = VOPAsmPrefer32Bit;
let TSFlags{42} = VOP3_OPSEL;
let TSFlags{43} = maybeAtomic;
let TSFlags{44} = renamedInGFX9;
let TSFlags{45} = FPClamp;
let TSFlags{46} = IntClamp;
let TSFlags{47} = ClampLo;
let TSFlags{48} = ClampHi;
let TSFlags{49} = IsPacked;
let TSFlags{50} = D16Buf;
let TSFlags{51} = FlatGlobal;
let TSFlags{52} = FPDPRounding;
let TSFlags{53} = FPAtomic;
let TSFlags{54} = IsMAI;
let TSFlags{55} = IsDOT;
let TSFlags{56} = FlatScratch;
let TSFlags{57} = IsAtomicNoRet;
let TSFlags{58} = IsAtomicRet;
let TSFlags{59} = IsWMMA;
let TSFlags{60} = TiedSourceNotRead;
let SchedRW = [Write32Bit];
let AsmVariantName = AMDGPUAsmVariants.Default;
// Avoid changing source registers in a way that violates constant bus read limitations.
let hasExtraSrcRegAllocReq = !or(VOP1, VOP2, VOP3, VOPC, SDWA, VALU);
}
class PseudoInstSI<dag outs, dag ins, list<dag> pattern = [], string asm = "">
: InstSI<outs, ins, asm, pattern> {
let isPseudo = 1;
let isCodeGenOnly = 1;
}
class SPseudoInstSI<dag outs, dag ins, list<dag> pattern = [], string asm = "">
: PseudoInstSI<outs, ins, pattern, asm> {
let SALU = 1;
}
class VPseudoInstSI<dag outs, dag ins, list<dag> pattern = [], string asm = "">
: PseudoInstSI<outs, ins, pattern, asm> {
let VALU = 1;
let Uses = [EXEC];
}
class CFPseudoInstSI<dag outs, dag ins, list<dag> pattern = [],
bit UseExec = 0, bit DefExec = 0> :
SPseudoInstSI<outs, ins, pattern> {
let Uses = !if(UseExec, [EXEC], []);
let Defs = !if(DefExec, [EXEC, SCC], [SCC]);
let mayLoad = 0;
let mayStore = 0;
let hasSideEffects = 0;
}
class Enc32 {
field bits<32> Inst;
int Size = 4;
}
class Enc64 {
field bits<64> Inst;
int Size = 8;
}
class Enc96 {
field bits<96> Inst;
int Size = 12;
}
def CPolBit {
int GLC = 0;
int SLC = 1;
int DLC = 2;
int SCC = 4;
}
class VOPDstOperand <RegisterClass rc> : RegisterOperand <rc, "printVOPDst">;
class VINTRPe <bits<2> op> : Enc32 {
bits<8> vdst;
bits<8> vsrc;
bits<2> attrchan;
bits<6> attr;
let Inst{7-0} = vsrc;
let Inst{9-8} = attrchan;
let Inst{15-10} = attr;
let Inst{17-16} = op;
let Inst{25-18} = vdst;
let Inst{31-26} = 0x32; // encoding
}
class MIMGe_gfxpre11 : Enc64 {
bits<10> vdata;
bits<4> dmask;
bits<1> unorm;
bits<5> cpol;
bits<1> r128;
bits<1> tfe;
bits<1> lwe;
bit d16;
bits<7> srsrc;
bits<7> ssamp;
let Inst{11-8} = dmask;
let Inst{12} = unorm;
let Inst{13} = cpol{CPolBit.GLC};
let Inst{15} = r128;
let Inst{17} = lwe;
let Inst{25} = cpol{CPolBit.SLC};
let Inst{31-26} = 0x3c;
let Inst{47-40} = vdata{7-0};
let Inst{52-48} = srsrc{6-2};
let Inst{57-53} = ssamp{6-2};
let Inst{63} = d16;
}
class MIMGe_gfx6789 <bits<8> op> : MIMGe_gfxpre11 {
bits<8> vaddr;
bits<1> da;
let Inst{0} = op{7};
let Inst{7} = cpol{CPolBit.SCC};
let Inst{14} = da;
let Inst{16} = tfe;
let Inst{24-18} = op{6-0};
let Inst{39-32} = vaddr;
}
class MIMGe_gfx90a <bits<8> op> : MIMGe_gfxpre11 {
bits<8> vaddr;
bits<1> da;
let Inst{0} = op{7};
let Inst{7} = cpol{CPolBit.SCC};
let Inst{14} = da;
let Inst{16} = vdata{9}; // ACC bit
let Inst{24-18} = op{6-0};
let Inst{39-32} = vaddr;
}
class MIMGe_gfx10 <bits<8> op> : MIMGe_gfxpre11 {
bits<8> vaddr0;
bits<3> dim;
bits<2> nsa;
bits<1> a16;
let Inst{0} = op{7};
let Inst{2-1} = nsa;
let Inst{5-3} = dim;
let Inst{7} = cpol{CPolBit.DLC};
let Inst{16} = tfe;
let Inst{24-18} = op{6-0};
let Inst{39-32} = vaddr0;
let Inst{62} = a16;
}
class MIMGe_gfx11 <bits<8> op> : Enc64 {
bits<8> vdata;
bits<4> dmask;
bits<1> unorm;
bits<5> cpol;
bits<1> r128;
bits<1> tfe;
bits<1> lwe;
bits<7> srsrc;
bits<7> ssamp;
bit d16;
bits<1> a16;
bits<8> vaddr0;
bits<3> dim;
bits<1> nsa;
let Inst{0} = nsa;
let Inst{4-2} = dim;
let Inst{7} = unorm;
let Inst{11-8} = dmask;
let Inst{12} = cpol{CPolBit.SLC};
let Inst{13} = cpol{CPolBit.DLC};
let Inst{14} = cpol{CPolBit.GLC};
let Inst{15} = r128;
let Inst{16} = a16;
let Inst{17} = d16;
let Inst{25-18} = op;
let Inst{31-26} = 0x3c;
let Inst{39-32} = vaddr0;
let Inst{47-40} = vdata;
let Inst{52-48} = srsrc{6-2};
let Inst{53} = tfe;
let Inst{54} = lwe;
let Inst{62-58} = ssamp{6-2};
}
class EXPe : Enc64 {
bits<4> en;
bits<6> tgt;
bits<1> done;
bits<8> src0;
bits<8> src1;
bits<8> src2;
bits<8> src3;
let Inst{3-0} = en;
let Inst{9-4} = tgt;
let Inst{11} = done;
let Inst{31-26} = 0x3e;
let Inst{39-32} = src0;
let Inst{47-40} = src1;
let Inst{55-48} = src2;
let Inst{63-56} = src3;
}
// Pre-GFX11 encoding has compr and vm bits.
class EXPe_ComprVM : EXPe {
bits<1> compr;
bits<1> vm;
let Inst{10} = compr;
let Inst{12} = vm;
}
// GFX11+ encoding has row bit.
class EXPe_Row : EXPe {
bits<1> row;
let Inst{13} = row;
}
let Uses = [EXEC] in {
class VINTRPCommon <dag outs, dag ins, string asm, list<dag> pattern> :
InstSI <outs, ins, asm, pattern> {
let VINTRP = 1;
// VINTRP instructions read parameter values from LDS, but these parameter
// values are stored outside of the LDS memory that is allocated to the
// shader for general purpose use.
//
// While it may be possible for ds_read/ds_write instructions to access
// the parameter values in LDS, this would essentially be an out-of-bounds
// memory access which we consider to be undefined behavior.
//
// So even though these instructions read memory, this memory is outside the
// addressable memory space for the shader, and we consider these instructions
// to be readnone.
let mayLoad = 0;
let mayStore = 0;
let hasSideEffects = 0;
let VALU = 1;
}
} // End Uses = [EXEC]