xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/EvergreenInstructions.td (revision 349cc55c9796c4596a5b9904cd3281af295f878f)

//===-- EvergreenInstructions.td - EG Instruction defs  ----*- tablegen -*-===//
//
// 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
//
//===----------------------------------------------------------------------===//
//
// TableGen definitions for instructions which are:
// - Available to Evergreen and newer VLIW4/VLIW5 GPUs
// - Available only on Evergreen family GPUs.
//
//===----------------------------------------------------------------------===//

def isEG : Predicate<
  "Subtarget->getGeneration() >= AMDGPUSubtarget::EVERGREEN && "
  "!Subtarget->hasCaymanISA()"
>;

def isEGorCayman : Predicate<
  "Subtarget->getGeneration() == AMDGPUSubtarget::EVERGREEN ||"
  "Subtarget->getGeneration() == AMDGPUSubtarget::NORTHERN_ISLANDS"
>;

class EGPat<dag pattern, dag result> : AMDGPUPat<pattern, result> {
  let SubtargetPredicate = isEG;
}

class EGOrCaymanPat<dag pattern, dag result> : AMDGPUPat<pattern, result> {
  let SubtargetPredicate = isEGorCayman;
}

def IMMZeroBasedBitfieldMask : ImmLeaf <i32, [{
  return isMask_32(Imm);
}]>;

def IMMPopCount : SDNodeXForm<imm, [{
  return CurDAG->getTargetConstant(countPopulation(N->getZExtValue()), SDLoc(N),
                                   MVT::i32);
}]>;

//===----------------------------------------------------------------------===//
// Evergreen / Cayman store instructions
//===----------------------------------------------------------------------===//

let SubtargetPredicate = isEGorCayman in {

class CF_MEM_RAT_CACHELESS <bits<6> rat_inst, bits<4> rat_id, bits<4> mask, dag ins,
                           string name, list<dag> pattern>
    : EG_CF_RAT <0x57, rat_inst, rat_id, mask, (outs), ins,
                 "MEM_RAT_CACHELESS "#name, pattern>;

class CF_MEM_RAT <bits<6> rat_inst, bits<4> rat_id, bits<4> mask, dag ins,
                  dag outs, string name, list<dag> pattern>
    : EG_CF_RAT <0x56, rat_inst, rat_id, mask, outs, ins,
                 "MEM_RAT "#name, pattern>;

class CF_MEM_RAT_STORE_TYPED<bits<1> has_eop>
    : CF_MEM_RAT <0x1, ?, 0xf, (ins R600_Reg128:$rw_gpr, R600_Reg128:$index_gpr,
                           i32imm:$rat_id, InstFlag:$eop), (outs),
                  "STORE_TYPED RAT($rat_id) $rw_gpr, $index_gpr"
                               #!if(has_eop, ", $eop", ""),
                  [(int_r600_rat_store_typed R600_Reg128:$rw_gpr,
                                             R600_Reg128:$index_gpr,
                                             (i32 imm:$rat_id))]>;

def RAT_MSKOR : CF_MEM_RAT <0x11, 0, 0xf,
  (ins R600_Reg128:$rw_gpr, R600_TReg32_X:$index_gpr), (outs),
  "MSKOR $rw_gpr.XW, $index_gpr",
  [(mskor_global v4i32:$rw_gpr, i32:$index_gpr)]
> {
  let eop = 0;
}


multiclass RAT_ATOMIC<bits<6> op_ret, bits<6> op_noret, string name> {
  let Constraints = "$rw_gpr = $out_gpr", eop = 0, mayStore = 1 in {
  def  _RTN: CF_MEM_RAT <op_ret, 0, 0xf,
             (ins R600_Reg128:$rw_gpr, R600_TReg32_X:$index_gpr),
             (outs R600_Reg128:$out_gpr),
             name # "_RTN $rw_gpr, $index_gpr", [] >;
  def _NORET: CF_MEM_RAT <op_noret, 0, 0xf,
              (ins R600_Reg128:$rw_gpr, R600_TReg32_X:$index_gpr),
              (outs R600_Reg128:$out_gpr),
              name # " $rw_gpr, $index_gpr", [] >;
  }
}

// Swap no-ret is just store. Raw store to cached target
// can only store on dword, which exactly matches swap_no_ret.
defm RAT_ATOMIC_XCHG_INT : RAT_ATOMIC<1, 34, "ATOMIC_XCHG_INT">;
defm RAT_ATOMIC_CMPXCHG_INT : RAT_ATOMIC<4, 36, "ATOMIC_CMPXCHG_INT">;
defm RAT_ATOMIC_ADD : RAT_ATOMIC<7, 39, "ATOMIC_ADD">;
defm RAT_ATOMIC_SUB : RAT_ATOMIC<8, 40, "ATOMIC_SUB">;
defm RAT_ATOMIC_RSUB : RAT_ATOMIC<9, 41, "ATOMIC_RSUB">;
defm RAT_ATOMIC_MIN_INT : RAT_ATOMIC<10, 42, "ATOMIC_MIN_INT">;
defm RAT_ATOMIC_MIN_UINT : RAT_ATOMIC<11, 43, "ATOMIC_MIN_UINT">;
defm RAT_ATOMIC_MAX_INT : RAT_ATOMIC<12, 44, "ATOMIC_MAX_INT">;
defm RAT_ATOMIC_MAX_UINT : RAT_ATOMIC<13, 45, "ATOMIC_MAX_UINT">;
defm RAT_ATOMIC_AND : RAT_ATOMIC<14, 46, "ATOMIC_AND">;
defm RAT_ATOMIC_OR : RAT_ATOMIC<15, 47, "ATOMIC_OR">;
defm RAT_ATOMIC_XOR : RAT_ATOMIC<16, 48, "ATOMIC_XOR">;
defm RAT_ATOMIC_INC_UINT : RAT_ATOMIC<18, 50, "ATOMIC_INC_UINT">;
defm RAT_ATOMIC_DEC_UINT : RAT_ATOMIC<19, 51, "ATOMIC_DEC_UINT">;

} // End SubtargetPredicate = isEGorCayman

//===----------------------------------------------------------------------===//
// Evergreen Only instructions
//===----------------------------------------------------------------------===//

let SubtargetPredicate = isEG in {

def RECIP_IEEE_eg : RECIP_IEEE_Common<0x86>;
defm DIV_eg : DIV_Common<RECIP_IEEE_eg>;

def MULLO_INT_eg : MULLO_INT_Common<0x8F>;
def MULHI_INT_eg : MULHI_INT_Common<0x90>;
def MULLO_UINT_eg : MULLO_UINT_Common<0x91>;
def MULHI_UINT_eg : MULHI_UINT_Common<0x92>;
def MULHI_UINT24_eg : MULHI_UINT24_Common<0xb2>;

def RECIP_UINT_eg : RECIP_UINT_Common<0x94>;
def RECIPSQRT_CLAMPED_eg : RECIPSQRT_CLAMPED_Common<0x87>;
def EXP_IEEE_eg : EXP_IEEE_Common<0x81>;
def LOG_IEEE_eg : LOG_IEEE_Common<0x83>;
def RECIP_CLAMPED_eg : RECIP_CLAMPED_Common<0x84>;
def RECIPSQRT_IEEE_eg : RECIPSQRT_IEEE_Common<0x89>;
def : RsqPat<RECIPSQRT_IEEE_eg, f32>;
def : SqrtPat<RECIPSQRT_IEEE_eg, RECIP_IEEE_eg>;

def SIN_eg : SIN_Common<0x8D>;
def COS_eg : COS_Common<0x8E>;

def : POW_Common <LOG_IEEE_eg, EXP_IEEE_eg, MUL>;
} // End SubtargetPredicate = isEG

//===----------------------------------------------------------------------===//
// Memory read/write instructions
//===----------------------------------------------------------------------===//

let usesCustomInserter = 1 in {

// 32-bit store
def RAT_WRITE_CACHELESS_32_eg : CF_MEM_RAT_CACHELESS <0x2, 0, 0x1,
  (ins R600_TReg32_X:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
  "STORE_RAW $rw_gpr, $index_gpr, $eop",
  [(store_global i32:$rw_gpr, i32:$index_gpr)]
>;

// 64-bit store
def RAT_WRITE_CACHELESS_64_eg : CF_MEM_RAT_CACHELESS <0x2, 0, 0x3,
  (ins R600_Reg64:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
  "STORE_RAW $rw_gpr.XY, $index_gpr, $eop",
  [(store_global v2i32:$rw_gpr, i32:$index_gpr)]
>;

//128-bit store
def RAT_WRITE_CACHELESS_128_eg : CF_MEM_RAT_CACHELESS <0x2, 0, 0xf,
  (ins R600_Reg128:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
  "STORE_RAW $rw_gpr.XYZW, $index_gpr, $eop",
  [(store_global v4i32:$rw_gpr, i32:$index_gpr)]
>;

def RAT_STORE_TYPED_eg: CF_MEM_RAT_STORE_TYPED<1>;

} // End usesCustomInserter = 1

class VTX_READ_eg <string name, dag outs>
    : VTX_WORD0_eg, VTX_READ<name, outs, []> {

  // Static fields
  let VC_INST = 0;
  let FETCH_TYPE = 2;
  let FETCH_WHOLE_QUAD = 0;
  let SRC_REL = 0;
  // XXX: We can infer this field based on the SRC_GPR.  This would allow us
  // to store vertex addresses in any channel, not just X.
  let SRC_SEL_X = 0;

  let Inst{31-0} = Word0;
}

def VTX_READ_8_eg
    : VTX_READ_eg <"VTX_READ_8 $dst_gpr, $src_gpr",
                   (outs R600_TReg32_X:$dst_gpr)> {

  let MEGA_FETCH_COUNT = 1;
  let DST_SEL_X = 0;
  let DST_SEL_Y = 7;   // Masked
  let DST_SEL_Z = 7;   // Masked
  let DST_SEL_W = 7;   // Masked
  let DATA_FORMAT = 1; // FMT_8
}

def VTX_READ_16_eg
    : VTX_READ_eg <"VTX_READ_16 $dst_gpr, $src_gpr",
                   (outs R600_TReg32_X:$dst_gpr)> {
  let MEGA_FETCH_COUNT = 2;
  let DST_SEL_X = 0;
  let DST_SEL_Y = 7;   // Masked
  let DST_SEL_Z = 7;   // Masked
  let DST_SEL_W = 7;   // Masked
  let DATA_FORMAT = 5; // FMT_16

}

def VTX_READ_32_eg
    : VTX_READ_eg <"VTX_READ_32 $dst_gpr, $src_gpr",
                   (outs R600_TReg32_X:$dst_gpr)> {

  let MEGA_FETCH_COUNT = 4;
  let DST_SEL_X        = 0;
  let DST_SEL_Y        = 7;   // Masked
  let DST_SEL_Z        = 7;   // Masked
  let DST_SEL_W        = 7;   // Masked
  let DATA_FORMAT      = 0xD; // COLOR_32

  // This is not really necessary, but there were some GPU hangs that appeared
  // to be caused by ALU instructions in the next instruction group that wrote
  // to the $src_gpr registers of the VTX_READ.
  // e.g.
  // %t3_x = VTX_READ_PARAM_32_eg killed %t2_x, 24
  // %t2_x = MOV %zero
  //Adding this constraint prevents this from happening.
  let Constraints = "$src_gpr.ptr = $dst_gpr";
}

def VTX_READ_64_eg
    : VTX_READ_eg <"VTX_READ_64 $dst_gpr.XY, $src_gpr",
                   (outs R600_Reg64:$dst_gpr)> {

  let MEGA_FETCH_COUNT = 8;
  let DST_SEL_X        = 0;
  let DST_SEL_Y        = 1;
  let DST_SEL_Z        = 7;
  let DST_SEL_W        = 7;
  let DATA_FORMAT      = 0x1D; // COLOR_32_32
}

def VTX_READ_128_eg
    : VTX_READ_eg <"VTX_READ_128 $dst_gpr.XYZW, $src_gpr",
                   (outs R600_Reg128:$dst_gpr)> {

  let MEGA_FETCH_COUNT = 16;
  let DST_SEL_X        =  0;
  let DST_SEL_Y        =  1;
  let DST_SEL_Z        =  2;
  let DST_SEL_W        =  3;
  let DATA_FORMAT      =  0x22; // COLOR_32_32_32_32

  // XXX: Need to force VTX_READ_128 instructions to write to the same register
  // that holds its buffer address to avoid potential hangs.  We can't use
  // the same constraint as VTX_READ_32_eg, because the $src_gpr.ptr and $dst
  // registers are different sizes.
}

//===----------------------------------------------------------------------===//
// VTX Read from parameter memory space
//===----------------------------------------------------------------------===//
def : EGPat<(i32:$dst_gpr (vtx_id3_az_extloadi8 ADDRVTX_READ:$src_gpr)),
          (VTX_READ_8_eg MEMxi:$src_gpr, 3)>;
def : EGPat<(i32:$dst_gpr (vtx_id3_az_extloadi16 ADDRVTX_READ:$src_gpr)),
          (VTX_READ_16_eg MEMxi:$src_gpr, 3)>;
def : EGPat<(i32:$dst_gpr (vtx_id3_load ADDRVTX_READ:$src_gpr)),
          (VTX_READ_32_eg MEMxi:$src_gpr, 3)>;
def : EGPat<(v2i32:$dst_gpr (vtx_id3_load ADDRVTX_READ:$src_gpr)),
          (VTX_READ_64_eg MEMxi:$src_gpr, 3)>;
def : EGPat<(v4i32:$dst_gpr (vtx_id3_load ADDRVTX_READ:$src_gpr)),
          (VTX_READ_128_eg MEMxi:$src_gpr, 3)>;

//===----------------------------------------------------------------------===//
// VTX Read from constant memory space
//===----------------------------------------------------------------------===//
def : EGPat<(i32:$dst_gpr (vtx_id2_az_extloadi8 ADDRVTX_READ:$src_gpr)),
          (VTX_READ_8_eg MEMxi:$src_gpr, 2)>;
def : EGPat<(i32:$dst_gpr (vtx_id2_az_extloadi16 ADDRVTX_READ:$src_gpr)),
          (VTX_READ_16_eg MEMxi:$src_gpr, 2)>;
def : EGPat<(i32:$dst_gpr (vtx_id2_load ADDRVTX_READ:$src_gpr)),
          (VTX_READ_32_eg MEMxi:$src_gpr, 2)>;
def : EGPat<(v2i32:$dst_gpr (vtx_id2_load ADDRVTX_READ:$src_gpr)),
          (VTX_READ_64_eg MEMxi:$src_gpr, 2)>;
def : EGPat<(v4i32:$dst_gpr (vtx_id2_load ADDRVTX_READ:$src_gpr)),
          (VTX_READ_128_eg MEMxi:$src_gpr, 2)>;

//===----------------------------------------------------------------------===//
// VTX Read from global memory space
//===----------------------------------------------------------------------===//
def : EGPat<(i32:$dst_gpr (vtx_id1_az_extloadi8 ADDRVTX_READ:$src_gpr)),
          (VTX_READ_8_eg MEMxi:$src_gpr, 1)>;
def : EGPat<(i32:$dst_gpr (vtx_id1_az_extloadi16 ADDRVTX_READ:$src_gpr)),
          (VTX_READ_16_eg MEMxi:$src_gpr, 1)>;
def : EGPat<(i32:$dst_gpr (vtx_id1_load ADDRVTX_READ:$src_gpr)),
          (VTX_READ_32_eg MEMxi:$src_gpr, 1)>;
def : EGPat<(v2i32:$dst_gpr (vtx_id1_load ADDRVTX_READ:$src_gpr)),
          (VTX_READ_64_eg MEMxi:$src_gpr, 1)>;
def : EGPat<(v4i32:$dst_gpr (vtx_id1_load ADDRVTX_READ:$src_gpr)),
          (VTX_READ_128_eg MEMxi:$src_gpr, 1)>;

//===----------------------------------------------------------------------===//
// Evergreen / Cayman Instructions
//===----------------------------------------------------------------------===//

let SubtargetPredicate = isEGorCayman in {

multiclass AtomicPat<Instruction inst_noret,
                     SDPatternOperator node_noret> {
  // FIXME: Add _RTN version. We need per WI scratch location to store the old value
  // EXTRACT_SUBREG here is dummy, we know the node has no uses
  def : EGOrCaymanPat<(i32 (node_noret i32:$ptr, i32:$data)),
            (EXTRACT_SUBREG (inst_noret
              (INSERT_SUBREG (v4i32 (IMPLICIT_DEF)), $data, sub0), $ptr), sub1)>;
}
multiclass AtomicIncDecPat<Instruction inst_noret,
                           SDPatternOperator node_noret, int C> {
  // FIXME: Add _RTN version. We need per WI scratch location to store the old value
  // EXTRACT_SUBREG here is dummy, we know the node has no uses
  def : EGOrCaymanPat<(i32 (node_noret i32:$ptr, C)),
            (EXTRACT_SUBREG (inst_noret
              (INSERT_SUBREG (v4i32 (IMPLICIT_DEF)), (MOV_IMM_I32 -1), sub0), $ptr), sub1)>;
}

// CMPSWAP is pattern is special
// EXTRACT_SUBREG here is dummy, we know the node has no uses
// FIXME: Add _RTN version. We need per WI scratch location to store the old value
def : EGOrCaymanPat<(i32 (atomic_cmp_swap_global_noret i32:$ptr, i32:$cmp, i32:$data)),
          (EXTRACT_SUBREG (RAT_ATOMIC_CMPXCHG_INT_NORET
            (INSERT_SUBREG
              (INSERT_SUBREG (v4i32 (IMPLICIT_DEF)), $cmp, sub3),
            $data, sub0),
          $ptr), sub1)>;

defm AtomicSwapPat : AtomicPat <RAT_ATOMIC_XCHG_INT_NORET,
                                atomic_swap_global_noret_32>;
defm AtomicAddPat : AtomicPat <RAT_ATOMIC_ADD_NORET,
                               atomic_load_add_global_noret_32>;
defm AtomicSubPat : AtomicPat <RAT_ATOMIC_SUB_NORET,
                               atomic_load_sub_global_noret_32>;
defm AtomicMinPat : AtomicPat <RAT_ATOMIC_MIN_INT_NORET,
                               atomic_load_min_global_noret_32>;
defm AtomicUMinPat : AtomicPat <RAT_ATOMIC_MIN_UINT_NORET,
                                atomic_load_umin_global_noret_32>;
defm AtomicMaxPat : AtomicPat <RAT_ATOMIC_MAX_INT_NORET,
                               atomic_load_max_global_noret_32>;
defm AtomicUMaxPat : AtomicPat <RAT_ATOMIC_MAX_UINT_NORET,
                                atomic_load_umax_global_noret_32>;
defm AtomicAndPat : AtomicPat <RAT_ATOMIC_AND_NORET,
                               atomic_load_and_global_noret_32>;
defm AtomicOrPat : AtomicPat <RAT_ATOMIC_OR_NORET,
                              atomic_load_or_global_noret_32>;
defm AtomicXorPat : AtomicPat <RAT_ATOMIC_XOR_NORET,
                               atomic_load_xor_global_noret_32>;
defm AtomicIncAddPat : AtomicIncDecPat <RAT_ATOMIC_INC_UINT_NORET,
                                        atomic_load_add_global_noret_32, 1>;
defm AtomicIncSubPat : AtomicIncDecPat <RAT_ATOMIC_INC_UINT_NORET,
                                        atomic_load_sub_global_noret_32, -1>;
defm AtomicDecAddPat : AtomicIncDecPat <RAT_ATOMIC_DEC_UINT_NORET,
                                        atomic_load_add_global_noret_32, -1>;
defm AtomicDecSubPat : AtomicIncDecPat <RAT_ATOMIC_DEC_UINT_NORET,
                                        atomic_load_sub_global_noret_32, 1>;

// Should be predicated on FeatureFP64
// def FMA_64 : R600_3OP <
//   0xA, "FMA_64",
//   [(set f64:$dst, (fma f64:$src0, f64:$src1, f64:$src2))]
// >;

// BFE_UINT - bit_extract, an optimization for mask and shift
// Src0 = Input
// Src1 = Offset
// Src2 = Width
//
// bit_extract = (Input << (32 - Offset - Width)) >> (32 - Width)
//
// Example Usage:
// (Offset, Width)
//
// (0, 8)  = (Input << 24) >> 24 = (Input &  0xff)       >> 0
// (8, 8)  = (Input << 16) >> 24 = (Input &  0xffff)     >> 8
// (16, 8) = (Input <<  8) >> 24 = (Input &  0xffffff)   >> 16
// (24, 8) = (Input <<  0) >> 24 = (Input &  0xffffffff) >> 24
def BFE_UINT_eg : R600_3OP <0x4, "BFE_UINT",
  [(set i32:$dst, (AMDGPUbfe_u32 i32:$src0, i32:$src1, i32:$src2))],
  VecALU
>;

def BFE_INT_eg : R600_3OP <0x5, "BFE_INT",
  [(set i32:$dst, (AMDGPUbfe_i32 i32:$src0, i32:$src1, i32:$src2))],
  VecALU
>;

// Bitfield extract patterns

def : AMDGPUPat <
  (and (i32 (srl i32:$src, i32:$rshift)), IMMZeroBasedBitfieldMask:$mask),
  (BFE_UINT_eg $src, $rshift, (MOV_IMM_I32 (i32 (IMMPopCount $mask))))
>;

// x & ((1 << y) - 1)
def : AMDGPUPat <
  (and i32:$src, (add_oneuse (shl_oneuse 1, i32:$width), -1)),
  (BFE_UINT_eg $src, (MOV_IMM_I32 (i32 0)), $width)
>;

// x & ~(-1 << y)
def : AMDGPUPat <
  (and i32:$src, (xor_oneuse (shl_oneuse -1, i32:$width), -1)),
  (BFE_UINT_eg $src, (MOV_IMM_I32 (i32 0)), $width)
>;

// x & (-1 >> (bitwidth - y))
def : AMDGPUPat <
  (and i32:$src, (srl_oneuse -1, (sub 32, i32:$width))),
  (BFE_UINT_eg $src, (MOV_IMM_I32 (i32 0)), $width)
>;

// x << (bitwidth - y) >> (bitwidth - y)
def : AMDGPUPat <
  (srl (shl_oneuse i32:$src, (sub 32, i32:$width)), (sub 32, i32:$width)),
  (BFE_UINT_eg $src, (MOV_IMM_I32 (i32 0)), $width)
>;

def : AMDGPUPat <
  (sra (shl_oneuse i32:$src, (sub 32, i32:$width)), (sub 32, i32:$width)),
  (BFE_INT_eg $src, (MOV_IMM_I32 (i32 0)), $width)
>;

def BFI_INT_eg : R600_3OP <0x06, "BFI_INT",
  [(set i32:$dst, (AMDGPUbfi i32:$src0, i32:$src1, i32:$src2))],
  VecALU
>;

def : EGOrCaymanPat<(i32 (sext_inreg i32:$src, i1)),
  (BFE_INT_eg i32:$src, (i32 ZERO), (i32 ONE_INT))>;
def : EGOrCaymanPat<(i32 (sext_inreg i32:$src, i8)),
  (BFE_INT_eg i32:$src, (i32 ZERO), (MOV_IMM_I32 8))>;
def : EGOrCaymanPat<(i32 (sext_inreg i32:$src, i16)),
  (BFE_INT_eg i32:$src, (i32 ZERO), (MOV_IMM_I32 16))>;

// BFI patterns

// Definition from ISA doc:
// (y & x) | (z & ~x)
def : AMDGPUPat <
  (or (and i32:$y, i32:$x), (and i32:$z, (not i32:$x))),
  (BFI_INT_eg $x, $y, $z)
>;

// 64-bit version
def : AMDGPUPat <
  (or (and i64:$y, i64:$x), (and i64:$z, (not i64:$x))),
  (REG_SEQUENCE R600_Reg64,
    (BFI_INT_eg (i32 (EXTRACT_SUBREG R600_Reg64:$x, sub0)),
                (i32 (EXTRACT_SUBREG R600_Reg64:$y, sub0)),
                (i32 (EXTRACT_SUBREG R600_Reg64:$z, sub0))), sub0,
    (BFI_INT_eg (i32 (EXTRACT_SUBREG R600_Reg64:$x, sub1)),
                (i32 (EXTRACT_SUBREG R600_Reg64:$y, sub1)),
                (i32 (EXTRACT_SUBREG R600_Reg64:$z, sub1))), sub1)
>;

// SHA-256 Ch function
// z ^ (x & (y ^ z))
def : AMDGPUPat <
  (xor i32:$z, (and i32:$x, (xor i32:$y, i32:$z))),
  (BFI_INT_eg $x, $y, $z)
>;

// 64-bit version
def : AMDGPUPat <
  (xor i64:$z, (and i64:$x, (xor i64:$y, i64:$z))),
  (REG_SEQUENCE R600_Reg64,
    (BFI_INT_eg (i32 (EXTRACT_SUBREG R600_Reg64:$x, sub0)),
                (i32 (EXTRACT_SUBREG R600_Reg64:$y, sub0)),
                (i32 (EXTRACT_SUBREG R600_Reg64:$z, sub0))), sub0,
    (BFI_INT_eg (i32 (EXTRACT_SUBREG R600_Reg64:$x, sub1)),
                (i32 (EXTRACT_SUBREG R600_Reg64:$y, sub1)),
                (i32 (EXTRACT_SUBREG R600_Reg64:$z, sub1))), sub1)
>;

def : AMDGPUPat <
  (fcopysign f32:$src0, f32:$src1),
  (BFI_INT_eg (MOV_IMM_I32 (i32 0x7fffffff)), $src0, $src1)
>;

def : AMDGPUPat <
  (fcopysign f32:$src0, f64:$src1),
  (BFI_INT_eg (MOV_IMM_I32 (i32 0x7fffffff)), $src0,
              (i32 (EXTRACT_SUBREG R600_Reg64:$src1, sub1)))
>;

def : AMDGPUPat <
  (fcopysign f64:$src0, f64:$src1),
  (REG_SEQUENCE R600_Reg64,
    (i32 (EXTRACT_SUBREG $src0, sub0)), sub0,
    (BFI_INT_eg (MOV_IMM_I32 (i32 0x7fffffff)),
                (i32 (EXTRACT_SUBREG R600_Reg64:$src0, sub1)),
                (i32 (EXTRACT_SUBREG R600_Reg64:$src1, sub1))), sub1)
>;

def : AMDGPUPat <
  (fcopysign f64:$src0, f32:$src1),
  (REG_SEQUENCE R600_Reg64,
    (i32 (EXTRACT_SUBREG $src0, sub0)), sub0,
    (BFI_INT_eg (MOV_IMM_I32 (i32 0x7fffffff)),
                (i32 (EXTRACT_SUBREG R600_Reg64:$src0, sub1)),
                $src1), sub1)
>;

def BFM_INT_eg : R600_2OP <0xA0, "BFM_INT",
  [(set i32:$dst, (AMDGPUbfm i32:$src0, i32:$src1))],
  VecALU
>;

def MULADD_UINT24_eg : R600_3OP <0x10, "MULADD_UINT24",
  [(set i32:$dst, (AMDGPUmad_u24 i32:$src0, i32:$src1, i32:$src2))], VecALU
>;

def : UMad24Pat<MULADD_UINT24_eg>;

def BIT_ALIGN_INT_eg : R600_3OP <0xC, "BIT_ALIGN_INT", [], VecALU>;
def : AMDGPUPat <
  (fshr i32:$src0, i32:$src1, i32:$src2),
  (BIT_ALIGN_INT_eg $src0, $src1, $src2)
>;
def : ROTRPattern <BIT_ALIGN_INT_eg>;
def MULADD_eg : MULADD_Common<0x14>;
def MULADD_IEEE_eg : MULADD_IEEE_Common<0x18>;
def FMA_eg : FMA_Common<0x7>;
def ASHR_eg : ASHR_Common<0x15>;
def LSHR_eg : LSHR_Common<0x16>;
def LSHL_eg : LSHL_Common<0x17>;
def CNDE_eg : CNDE_Common<0x19>;
def CNDGT_eg : CNDGT_Common<0x1A>;
def CNDGE_eg : CNDGE_Common<0x1B>;
def MUL_LIT_eg : MUL_LIT_Common<0x1F>;
def LOG_CLAMPED_eg : LOG_CLAMPED_Common<0x82>;
def MUL_UINT24_eg : R600_2OP <0xB5, "MUL_UINT24",
  [(set i32:$dst, (AMDGPUmul_u24 i32:$src0, i32:$src1))], VecALU
>;
def DOT4_eg : DOT4_Common<0xBE>;
defm CUBE_eg : CUBE_Common<0xC0>;


def ADDC_UINT : R600_2OP_Helper <0x52, "ADDC_UINT", AMDGPUcarry>;
def SUBB_UINT : R600_2OP_Helper <0x53, "SUBB_UINT", AMDGPUborrow>;

def FLT32_TO_FLT16 : R600_1OP_Helper <0xA2, "FLT32_TO_FLT16", AMDGPUfp_to_f16, VecALU>;
def FLT16_TO_FLT32 : R600_1OP_Helper <0xA3, "FLT16_TO_FLT32", f16_to_fp, VecALU>;
def BCNT_INT : R600_1OP_Helper <0xAA, "BCNT_INT", ctpop, VecALU>;
def FFBH_UINT : R600_1OP_Helper <0xAB, "FFBH_UINT", AMDGPUffbh_u32, VecALU>;
def FFBL_INT : R600_1OP_Helper <0xAC, "FFBL_INT", AMDGPUffbl_b32, VecALU>;

let hasSideEffects = 1 in {
  def MOVA_INT_eg : R600_1OP <0xCC, "MOVA_INT", [], VecALU>;
}

def FLT_TO_INT_eg : FLT_TO_INT_Common<0x50> {
  let Pattern = [];
  let Itinerary = AnyALU;
}

def INT_TO_FLT_eg : INT_TO_FLT_Common<0x9B>;

def FLT_TO_UINT_eg : FLT_TO_UINT_Common<0x9A> {
  let Pattern = [];
}

def UINT_TO_FLT_eg : UINT_TO_FLT_Common<0x9C>;

def GROUP_BARRIER : InstR600 <
    (outs), (ins), "  GROUP_BARRIER", [(int_r600_group_barrier)], AnyALU>,
    R600ALU_Word0,
    R600ALU_Word1_OP2 <0x54> {

  let dst = 0;
  let dst_rel = 0;
  let src0 = 0;
  let src0_rel = 0;
  let src0_neg = 0;
  let src0_abs = 0;
  let src1 = 0;
  let src1_rel = 0;
  let src1_neg = 0;
  let src1_abs = 0;
  let write = 0;
  let omod = 0;
  let clamp = 0;
  let last = 1;
  let bank_swizzle = 0;
  let pred_sel = 0;
  let update_exec_mask = 0;
  let update_pred = 0;

  let Inst{31-0}  = Word0;
  let Inst{63-32} = Word1;

  let ALUInst = 1;
}

//===----------------------------------------------------------------------===//
// LDS Instructions
//===----------------------------------------------------------------------===//
class R600_LDS  <bits<6> op, dag outs, dag ins, string asm,
                 list<dag> pattern = []> :

    InstR600 <outs, ins, asm, pattern, XALU>,
    R600_ALU_LDS_Word0,
    R600LDS_Word1 {

  bits<6>  offset = 0;
  let lds_op = op;

  let Word1{27} = offset{0};
  let Word1{12} = offset{1};
  let Word1{28} = offset{2};
  let Word1{31} = offset{3};
  let Word0{12} = offset{4};
  let Word0{25} = offset{5};


  let Inst{31-0}  = Word0;
  let Inst{63-32} = Word1;

  let ALUInst = 1;
  let HasNativeOperands = 1;
  let UseNamedOperandTable = 1;
}

class R600_LDS_1A <bits<6> lds_op, string name, list<dag> pattern> : R600_LDS <
  lds_op,
  (outs R600_Reg32:$dst),
  (ins R600_Reg32:$src0, REL:$src0_rel, SEL:$src0_sel,
       LAST:$last, R600_Pred:$pred_sel,
       BANK_SWIZZLE:$bank_swizzle),
  "  "#name#" $last OQAP, $src0$src0_rel $pred_sel",
  pattern
  > {

  let src1 = 0;
  let src1_rel = 0;
  let src2 = 0;
  let src2_rel = 0;

  let usesCustomInserter = 1;
  let LDS_1A = 1;
  let DisableEncoding = "$dst";
}

class R600_LDS_1A1D <bits<6> lds_op, dag outs, string name, list<dag> pattern,
                     string dst =""> :
    R600_LDS <
  lds_op, outs,
  (ins R600_Reg32:$src0, REL:$src0_rel, SEL:$src0_sel,
       R600_Reg32:$src1, REL:$src1_rel, SEL:$src1_sel,
       LAST:$last, R600_Pred:$pred_sel,
       BANK_SWIZZLE:$bank_swizzle),
  "  "#name#" $last "#dst#"$src0$src0_rel, $src1$src1_rel, $pred_sel",
  pattern
  > {

  field string BaseOp;

  let src2 = 0;
  let src2_rel = 0;
  let LDS_1A1D = 1;
}

class R600_LDS_1A1D_NORET <bits<6> lds_op, string name, list<dag> pattern> :
    R600_LDS_1A1D <lds_op, (outs), name, pattern> {
  let BaseOp = name;
}

class R600_LDS_1A1D_RET <bits<6> lds_op, string name, list<dag> pattern> :
    R600_LDS_1A1D <lds_op,  (outs R600_Reg32:$dst), name#"_RET", pattern, "OQAP, "> {

  let BaseOp = name;
  let usesCustomInserter = 1;
  let DisableEncoding = "$dst";
}

class R600_LDS_1A2D <bits<6> lds_op, dag outs, string name, list<dag> pattern,
                     string dst =""> :
    R600_LDS <
  lds_op, outs,
  (ins R600_Reg32:$src0, REL:$src0_rel, SEL:$src0_sel,
       R600_Reg32:$src1, REL:$src1_rel, SEL:$src1_sel,
       R600_Reg32:$src2, REL:$src2_rel, SEL:$src2_sel,
       LAST:$last, R600_Pred:$pred_sel, BANK_SWIZZLE:$bank_swizzle),
  "  "#name# "$last "#dst#"$src0$src0_rel, $src1$src1_rel, $src2$src2_rel, $pred_sel",
  pattern> {

  field string BaseOp;

  let LDS_1A1D = 0;
  let LDS_1A2D = 1;
}

class R600_LDS_1A2D_NORET <bits<6> lds_op, string name, list<dag> pattern> :
    R600_LDS_1A2D <lds_op, (outs), name, pattern> {
  let BaseOp = name;
}

class R600_LDS_1A2D_RET <bits<6> lds_op, string name, list<dag> pattern> :
    R600_LDS_1A2D <lds_op, (outs R600_Reg32:$dst), name, pattern> {

  let BaseOp = name;
  let usesCustomInserter = 1;
  let DisableEncoding = "$dst";
}

def LDS_ADD : R600_LDS_1A1D_NORET <0x0, "LDS_ADD", [] >;
def LDS_SUB : R600_LDS_1A1D_NORET <0x1, "LDS_SUB", [] >;
def LDS_AND : R600_LDS_1A1D_NORET <0x9, "LDS_AND", [] >;
def LDS_OR : R600_LDS_1A1D_NORET <0xa, "LDS_OR", [] >;
def LDS_XOR : R600_LDS_1A1D_NORET <0xb, "LDS_XOR", [] >;
def LDS_WRXCHG: R600_LDS_1A1D_NORET <0xd, "LDS_WRXCHG", [] >;
def LDS_CMPST: R600_LDS_1A2D_NORET <0x10, "LDS_CMPST", [] >;
def LDS_MIN_INT : R600_LDS_1A1D_NORET <0x5, "LDS_MIN_INT", [] >;
def LDS_MAX_INT : R600_LDS_1A1D_NORET <0x6, "LDS_MAX_INT", [] >;
def LDS_MIN_UINT : R600_LDS_1A1D_NORET <0x7, "LDS_MIN_UINT", [] >;
def LDS_MAX_UINT : R600_LDS_1A1D_NORET <0x8, "LDS_MAX_UINT", [] >;
def LDS_WRITE : R600_LDS_1A1D_NORET <0xD, "LDS_WRITE",
  [(store_local (i32 R600_Reg32:$src1), R600_Reg32:$src0)]
>;
def LDS_BYTE_WRITE : R600_LDS_1A1D_NORET<0x12, "LDS_BYTE_WRITE",
  [(truncstorei8_local i32:$src1, i32:$src0)]
>;
def LDS_SHORT_WRITE : R600_LDS_1A1D_NORET<0x13, "LDS_SHORT_WRITE",
  [(truncstorei16_local i32:$src1, i32:$src0)]
>;
def LDS_ADD_RET : R600_LDS_1A1D_RET <0x20, "LDS_ADD",
  [(set i32:$dst, (atomic_load_add_local_32 i32:$src0, i32:$src1))]
>;
def LDS_SUB_RET : R600_LDS_1A1D_RET <0x21, "LDS_SUB",
  [(set i32:$dst, (atomic_load_sub_local_32 i32:$src0, i32:$src1))]
>;
def LDS_AND_RET : R600_LDS_1A1D_RET <0x29, "LDS_AND",
  [(set i32:$dst, (atomic_load_and_local_32 i32:$src0, i32:$src1))]
>;
def LDS_OR_RET : R600_LDS_1A1D_RET <0x2a, "LDS_OR",
  [(set i32:$dst, (atomic_load_or_local_32 i32:$src0, i32:$src1))]
>;
def LDS_XOR_RET : R600_LDS_1A1D_RET <0x2b, "LDS_XOR",
  [(set i32:$dst, (atomic_load_xor_local_32 i32:$src0, i32:$src1))]
>;
def LDS_MIN_INT_RET : R600_LDS_1A1D_RET <0x25, "LDS_MIN_INT",
  [(set i32:$dst, (atomic_load_min_local_32 i32:$src0, i32:$src1))]
>;
def LDS_MAX_INT_RET : R600_LDS_1A1D_RET <0x26, "LDS_MAX_INT",
  [(set i32:$dst, (atomic_load_max_local_32 i32:$src0, i32:$src1))]
>;
def LDS_MIN_UINT_RET : R600_LDS_1A1D_RET <0x27, "LDS_MIN_UINT",
  [(set i32:$dst, (atomic_load_umin_local_32 i32:$src0, i32:$src1))]
>;
def LDS_MAX_UINT_RET : R600_LDS_1A1D_RET <0x28, "LDS_MAX_UINT",
  [(set i32:$dst, (atomic_load_umax_local_32 i32:$src0, i32:$src1))]
>;
def LDS_WRXCHG_RET : R600_LDS_1A1D_RET <0x2d, "LDS_WRXCHG",
  [(set i32:$dst, (atomic_swap_local_32 i32:$src0, i32:$src1))]
>;
def LDS_CMPST_RET : R600_LDS_1A2D_RET <0x30, "LDS_CMPST",
  [(set i32:$dst, (atomic_cmp_swap_local_32 i32:$src0, i32:$src1, i32:$src2))]
>;
def LDS_READ_RET : R600_LDS_1A <0x32, "LDS_READ_RET",
  [(set (i32 R600_Reg32:$dst), (load_local R600_Reg32:$src0))]
>;
def LDS_BYTE_READ_RET : R600_LDS_1A <0x36, "LDS_BYTE_READ_RET",
  [(set i32:$dst, (sextloadi8_local i32:$src0))]
>;
def LDS_UBYTE_READ_RET : R600_LDS_1A <0x37, "LDS_UBYTE_READ_RET",
  [(set i32:$dst, (az_extloadi8_local i32:$src0))]
>;
def LDS_SHORT_READ_RET : R600_LDS_1A <0x38, "LDS_SHORT_READ_RET",
  [(set i32:$dst, (sextloadi16_local i32:$src0))]
>;
def LDS_USHORT_READ_RET : R600_LDS_1A <0x39, "LDS_USHORT_READ_RET",
  [(set i32:$dst, (az_extloadi16_local i32:$src0))]
>;

// TRUNC is used for the FLT_TO_INT instructions to work around a
// perceived problem where the rounding modes are applied differently
// depending on the instruction and the slot they are in.
// See:
// https://bugs.freedesktop.org/show_bug.cgi?id=50232
// Mesa commit: a1a0974401c467cb86ef818f22df67c21774a38c
//
// XXX: Lowering SELECT_CC will sometimes generate fp_to_[su]int nodes,
// which do not need to be truncated since the fp values are 0.0f or 1.0f.
// We should look into handling these cases separately.
def : EGOrCaymanPat<(fp_to_sint f32:$src0), (FLT_TO_INT_eg (TRUNC $src0))>;

def : EGOrCaymanPat<(fp_to_uint f32:$src0), (FLT_TO_UINT_eg (TRUNC $src0))>;

// SHA-256 Ma patterns

// ((x & z) | (y & (x | z))) -> BFI (XOR x, y), z, y
def : AMDGPUPat <
  (or (and i32:$x, i32:$z), (and i32:$y, (or i32:$x, i32:$z))),
  (BFI_INT_eg (XOR_INT i32:$x, i32:$y), i32:$z, i32:$y)
>;

def : AMDGPUPat <
  (or (and i64:$x, i64:$z), (and i64:$y, (or i64:$x, i64:$z))),
  (REG_SEQUENCE R600_Reg64,
    (BFI_INT_eg (XOR_INT (i32 (EXTRACT_SUBREG R600_Reg64:$x, sub0)),
                     (i32 (EXTRACT_SUBREG R600_Reg64:$y, sub0))),
                (i32 (EXTRACT_SUBREG R600_Reg64:$z, sub0)),
                (i32 (EXTRACT_SUBREG R600_Reg64:$y, sub0))), sub0,
    (BFI_INT_eg (XOR_INT (i32 (EXTRACT_SUBREG R600_Reg64:$x, sub1)),
                     (i32 (EXTRACT_SUBREG R600_Reg64:$y, sub1))),
                (i32 (EXTRACT_SUBREG R600_Reg64:$z, sub1)),
                (i32 (EXTRACT_SUBREG R600_Reg64:$y, sub1))), sub1)
>;

def EG_ExportSwz : ExportSwzInst {
  let Word1{19-16} = 0; // BURST_COUNT
  let Word1{20} = 0; // VALID_PIXEL_MODE
  let Word1{21} = eop;
  let Word1{29-22} = inst;
  let Word1{30} = 0; // MARK
  let Word1{31} = 1; // BARRIER
}
defm : ExportPattern<EG_ExportSwz, 83>;

def EG_ExportBuf : ExportBufInst {
  let Word1{19-16} = 0; // BURST_COUNT
  let Word1{20} = 0; // VALID_PIXEL_MODE
  let Word1{21} = eop;
  let Word1{29-22} = inst;
  let Word1{30} = 0; // MARK
  let Word1{31} = 1; // BARRIER
}
defm : SteamOutputExportPattern<EG_ExportBuf, 0x40, 0x41, 0x42, 0x43>;

def CF_TC_EG : CF_CLAUSE_EG<1, (ins i32imm:$ADDR, i32imm:$COUNT),
  "TEX $COUNT @$ADDR"> {
  let POP_COUNT = 0;
}
def CF_VC_EG : CF_CLAUSE_EG<2, (ins i32imm:$ADDR, i32imm:$COUNT),
  "VTX $COUNT @$ADDR"> {
  let POP_COUNT = 0;
}
def WHILE_LOOP_EG : CF_CLAUSE_EG<6, (ins i32imm:$ADDR),
  "LOOP_START_DX10 @$ADDR"> {
  let POP_COUNT = 0;
  let COUNT = 0;
}
def END_LOOP_EG : CF_CLAUSE_EG<5, (ins i32imm:$ADDR), "END_LOOP @$ADDR"> {
  let POP_COUNT = 0;
  let COUNT = 0;
}
def LOOP_BREAK_EG : CF_CLAUSE_EG<9, (ins i32imm:$ADDR),
  "LOOP_BREAK @$ADDR"> {
  let POP_COUNT = 0;
  let COUNT = 0;
}
def CF_CONTINUE_EG : CF_CLAUSE_EG<8, (ins i32imm:$ADDR),
  "CONTINUE @$ADDR"> {
  let POP_COUNT = 0;
  let COUNT = 0;
}
def CF_JUMP_EG : CF_CLAUSE_EG<10, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
  "JUMP @$ADDR POP:$POP_COUNT"> {
  let COUNT = 0;
}
def CF_PUSH_EG : CF_CLAUSE_EG<11, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
                              "PUSH @$ADDR POP:$POP_COUNT"> {
  let COUNT = 0;
}
def CF_ELSE_EG : CF_CLAUSE_EG<13, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
  "ELSE @$ADDR POP:$POP_COUNT"> {
  let COUNT = 0;
}
def CF_CALL_FS_EG : CF_CLAUSE_EG<19, (ins), "CALL_FS"> {
  let ADDR = 0;
  let COUNT = 0;
  let POP_COUNT = 0;
}
def POP_EG : CF_CLAUSE_EG<14, (ins i32imm:$ADDR, i32imm:$POP_COUNT),
  "POP @$ADDR POP:$POP_COUNT"> {
  let COUNT = 0;
}
def CF_END_EG :  CF_CLAUSE_EG<0, (ins), "CF_END"> {
  let COUNT = 0;
  let POP_COUNT = 0;
  let ADDR = 0;
  let END_OF_PROGRAM = 1;
}

} // End Predicates = [isEGorCayman]