You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

1071 lines
48 KiB

//===---- SMInstructions.td - Scalar Memory 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
//
//===----------------------------------------------------------------------===//
def smrd_offset_8 : NamedOperandU32<"SMRDOffset8",
NamedMatchClass<"SMRDOffset8">> {
let OperandType = "OPERAND_IMMEDIATE";
}
def smem_offset : NamedOperandU32<"SMEMOffset",
NamedMatchClass<"SMEMOffset">> {
let OperandType = "OPERAND_IMMEDIATE";
let EncoderMethod = "getSMEMOffsetEncoding";
let DecoderMethod = "decodeSMEMOffset";
}
//===----------------------------------------------------------------------===//
// Scalar Memory classes
//===----------------------------------------------------------------------===//
class SM_Pseudo <string opName, dag outs, dag ins, string asmOps, list<dag> pattern=[]> :
InstSI <outs, ins, "", pattern>,
SIMCInstr<opName, SIEncodingFamily.NONE> {
let isPseudo = 1;
let isCodeGenOnly = 1;
let LGKM_CNT = 1;
let SMRD = 1;
let mayStore = 0;
let mayLoad = 1;
let hasSideEffects = 0;
let UseNamedOperandTable = 1;
let SchedRW = [WriteSMEM];
string Mnemonic = opName;
string AsmOperands = asmOps;
bits<1> has_sbase = 1;
bits<1> has_sdst = 1;
bit has_glc = 0;
bit has_dlc = 0;
bits<1> has_offset = 1;
bits<1> offset_is_imm = 0;
bit is_buffer = 0;
}
class SM_Real <SM_Pseudo ps>
: InstSI<ps.OutOperandList, ps.InOperandList, ps.Mnemonic # ps.AsmOperands, []> {
let isPseudo = 0;
let isCodeGenOnly = 0;
Instruction Opcode = !cast<Instruction>(NAME);
// copy relevant pseudo op flags
let SubtargetPredicate = ps.SubtargetPredicate;
let AsmMatchConverter = ps.AsmMatchConverter;
let UseNamedOperandTable = ps.UseNamedOperandTable;
let SMRD = ps.SMRD;
bit is_buffer = ps.is_buffer;
// encoding
bits<7> sbase;
bits<7> sdst;
bits<32> offset;
bits<1> imm = !if(ps.has_offset, ps.offset_is_imm, 0);
}
class SM_Probe_Pseudo <string opName, dag ins, bit isImm>
: SM_Pseudo<opName, (outs), ins, " $sdata, $sbase, $offset"> {
let mayLoad = 0;
let mayStore = 0;
let has_glc = 0;
let LGKM_CNT = 0;
let ScalarStore = 0;
let hasSideEffects = 1;
let offset_is_imm = isImm;
let PseudoInstr = opName # !if(isImm, "_IMM", "_SGPR");
}
class SM_Load_Pseudo <string opName, dag outs, dag ins, string asmOps, list<dag> pattern=[]>
: SM_Pseudo<opName, outs, ins, asmOps, pattern> {
RegisterClass BaseClass;
let mayLoad = 1;
let mayStore = 0;
let has_glc = 1;
let has_dlc = 1;
}
class SM_Store_Pseudo <string opName, dag ins, string asmOps, list<dag> pattern = []>
: SM_Pseudo<opName, (outs), ins, asmOps, pattern> {
RegisterClass BaseClass;
RegisterClass SrcClass;
let mayLoad = 0;
let mayStore = 1;
let has_glc = 1;
let has_dlc = 1;
let ScalarStore = 1;
}
class SM_Discard_Pseudo <string opName, dag ins, bit isImm>
: SM_Pseudo<opName, (outs), ins, " $sbase, $offset"> {
let mayLoad = 0;
let mayStore = 0;
let has_glc = 0;
let has_sdst = 0;
let ScalarStore = 0;
let hasSideEffects = 1;
let offset_is_imm = isImm;
let PseudoInstr = opName # !if(isImm, "_IMM", "_SGPR");
}
multiclass SM_Pseudo_Loads<string opName,
RegisterClass baseClass,
RegisterClass dstClass> {
def _IMM : SM_Load_Pseudo <opName,
(outs dstClass:$sdst),
(ins baseClass:$sbase, i32imm:$offset, i1imm:$glc, i1imm:$dlc),
" $sdst, $sbase, $offset$glc$dlc", []> {
let offset_is_imm = 1;
let BaseClass = baseClass;
let PseudoInstr = opName # "_IMM";
let has_glc = 1;
let has_dlc = 1;
}
def _SGPR : SM_Load_Pseudo <opName,
(outs dstClass:$sdst),
(ins baseClass:$sbase, SReg_32:$soff, i1imm:$glc, i1imm:$dlc),
" $sdst, $sbase, $offset$glc$dlc", []> {
let BaseClass = baseClass;
let PseudoInstr = opName # "_SGPR";
let has_glc = 1;
let has_dlc = 1;
}
}
multiclass SM_Pseudo_Stores<string opName,
RegisterClass baseClass,
RegisterClass srcClass> {
def _IMM : SM_Store_Pseudo <opName,
(ins srcClass:$sdata, baseClass:$sbase, i32imm:$offset, i1imm:$glc, i1imm:$dlc),
" $sdata, $sbase, $offset$glc$dlc", []> {
let offset_is_imm = 1;
let BaseClass = baseClass;
let SrcClass = srcClass;
let PseudoInstr = opName # "_IMM";
}
def _SGPR : SM_Store_Pseudo <opName,
(ins srcClass:$sdata, baseClass:$sbase, SReg_32:$soff, i1imm:$glc, i1imm:$dlc),
" $sdata, $sbase, $offset$glc$dlc", []> {
let BaseClass = baseClass;
let SrcClass = srcClass;
let PseudoInstr = opName # "_SGPR";
}
}
multiclass SM_Pseudo_Discards<string opName> {
def _IMM : SM_Discard_Pseudo <opName, (ins SReg_64:$sbase, smem_offset:$offset), 1>;
def _SGPR : SM_Discard_Pseudo <opName, (ins SReg_64:$sbase, SReg_32:$offset), 0>;
}
class SM_Time_Pseudo<string opName, SDPatternOperator node = null_frag> : SM_Pseudo<
opName, (outs SReg_64_XEXEC:$sdst), (ins),
" $sdst", [(set i64:$sdst, (node))]> {
let hasSideEffects = 1;
// FIXME: This should be definitively mayStore = 0. TableGen
// brokenly tries to infer these based on the intrinsic properties
// corresponding to the IR attributes. The target intrinsics are
// considered as writing to memory for IR dependency purposes, but
// those can be modeled with hasSideEffects here. These also end up
// inferring differently for llvm.readcyclecounter and the amdgcn
// intrinsics.
let mayStore = ?;
let mayLoad = 1;
let has_sbase = 0;
let has_offset = 0;
}
class SM_Inval_Pseudo <string opName, SDPatternOperator node = null_frag> : SM_Pseudo<
opName, (outs), (ins), "", [(node)]> {
let hasSideEffects = 1;
let mayStore = 0;
let has_sdst = 0;
let has_sbase = 0;
let has_offset = 0;
}
multiclass SM_Pseudo_Probe<string opName, RegisterClass baseClass> {
def _IMM : SM_Probe_Pseudo <opName, (ins i8imm:$sdata, baseClass:$sbase, smem_offset:$offset), 1>;
def _SGPR : SM_Probe_Pseudo <opName, (ins i8imm:$sdata, baseClass:$sbase, SReg_32:$offset), 0>;
}
class SM_WaveId_Pseudo<string opName, SDPatternOperator node> : SM_Pseudo<
opName, (outs SReg_32_XM0_XEXEC:$sdst), (ins),
" $sdst", [(set i32:$sdst, (node))]> {
let hasSideEffects = 1;
let mayStore = 0;
let mayLoad = 1;
let has_sbase = 0;
let has_offset = 0;
}
//===----------------------------------------------------------------------===//
// Scalar Atomic Memory Classes
//===----------------------------------------------------------------------===//
class SM_Atomic_Pseudo <string opName,
dag outs, dag ins, string asmOps, bit isRet>
: SM_Pseudo<opName, outs, ins, asmOps, []> {
bit glc = isRet;
let mayLoad = 1;
let mayStore = 1;
let has_glc = 1;
let has_dlc = 1;
// Should these be set?
let ScalarStore = 1;
let hasSideEffects = 1;
let maybeAtomic = 1;
}
class SM_Pseudo_Atomic<string opName,
RegisterClass baseClass,
RegisterClass dataClass,
bit isImm,
bit isRet> :
SM_Atomic_Pseudo<opName,
!if(isRet, (outs dataClass:$sdst), (outs)),
!if(isImm,
(ins dataClass:$sdata, baseClass:$sbase, smem_offset:$offset, DLC:$dlc),
(ins dataClass:$sdata, baseClass:$sbase, SReg_32:$offset, DLC:$dlc)),
!if(isRet, " $sdst", " $sdata") # ", $sbase, $offset" # !if(isRet, " glc", "") # "$dlc",
isRet> {
let offset_is_imm = isImm;
let PseudoInstr = opName # !if(isImm,
!if(isRet, "_IMM_RTN", "_IMM"),
!if(isRet, "_SGPR_RTN", "_SGPR"));
let Constraints = !if(isRet, "$sdst = $sdata", "");
let DisableEncoding = !if(isRet, "$sdata", "");
}
multiclass SM_Pseudo_Atomics<string opName,
RegisterClass baseClass,
RegisterClass dataClass> {
def _IMM : SM_Pseudo_Atomic <opName, baseClass, dataClass, 1, 0>;
def _SGPR : SM_Pseudo_Atomic <opName, baseClass, dataClass, 0, 0>;
def _IMM_RTN : SM_Pseudo_Atomic <opName, baseClass, dataClass, 1, 1>;
def _SGPR_RTN : SM_Pseudo_Atomic <opName, baseClass, dataClass, 0, 1>;
}
//===----------------------------------------------------------------------===//
// Scalar Memory Instructions
//===----------------------------------------------------------------------===//
// We are using the SReg_32_XM0 and not the SReg_32 register class for 32-bit
// SMRD instructions, because the SReg_32_XM0 register class does not include M0
// and writing to M0 from an SMRD instruction will hang the GPU.
// XXX - SMEM instructions do not allow exec for data operand, but
// does sdst for SMRD on SI/CI?
defm S_LOAD_DWORD : SM_Pseudo_Loads <"s_load_dword", SReg_64, SReg_32_XM0_XEXEC>;
defm S_LOAD_DWORDX2 : SM_Pseudo_Loads <"s_load_dwordx2", SReg_64, SReg_64_XEXEC>;
defm S_LOAD_DWORDX4 : SM_Pseudo_Loads <"s_load_dwordx4", SReg_64, SReg_128>;
defm S_LOAD_DWORDX8 : SM_Pseudo_Loads <"s_load_dwordx8", SReg_64, SReg_256>;
defm S_LOAD_DWORDX16 : SM_Pseudo_Loads <"s_load_dwordx16", SReg_64, SReg_512>;
let is_buffer = 1 in {
defm S_BUFFER_LOAD_DWORD : SM_Pseudo_Loads <
"s_buffer_load_dword", SReg_128, SReg_32_XM0_XEXEC
>;
// FIXME: exec_lo/exec_hi appear to be allowed for SMRD loads on
// SI/CI, bit disallowed for SMEM on VI.
defm S_BUFFER_LOAD_DWORDX2 : SM_Pseudo_Loads <
"s_buffer_load_dwordx2", SReg_128, SReg_64_XEXEC
>;
defm S_BUFFER_LOAD_DWORDX4 : SM_Pseudo_Loads <
"s_buffer_load_dwordx4", SReg_128, SReg_128
>;
defm S_BUFFER_LOAD_DWORDX8 : SM_Pseudo_Loads <
"s_buffer_load_dwordx8", SReg_128, SReg_256
>;
defm S_BUFFER_LOAD_DWORDX16 : SM_Pseudo_Loads <
"s_buffer_load_dwordx16", SReg_128, SReg_512
>;
}
let SubtargetPredicate = HasScalarStores in {
defm S_STORE_DWORD : SM_Pseudo_Stores <"s_store_dword", SReg_64, SReg_32_XM0_XEXEC>;
defm S_STORE_DWORDX2 : SM_Pseudo_Stores <"s_store_dwordx2", SReg_64, SReg_64_XEXEC>;
defm S_STORE_DWORDX4 : SM_Pseudo_Stores <"s_store_dwordx4", SReg_64, SReg_128>;
let is_buffer = 1 in {
defm S_BUFFER_STORE_DWORD : SM_Pseudo_Stores <
"s_buffer_store_dword", SReg_128, SReg_32_XM0_XEXEC
>;
defm S_BUFFER_STORE_DWORDX2 : SM_Pseudo_Stores <
"s_buffer_store_dwordx2", SReg_128, SReg_64_XEXEC
>;
defm S_BUFFER_STORE_DWORDX4 : SM_Pseudo_Stores <
"s_buffer_store_dwordx4", SReg_128, SReg_128
>;
}
} // End SubtargetPredicate = HasScalarStores
let SubtargetPredicate = HasSMemTimeInst in
def S_MEMTIME : SM_Time_Pseudo <"s_memtime", int_amdgcn_s_memtime>;
def S_DCACHE_INV : SM_Inval_Pseudo <"s_dcache_inv", int_amdgcn_s_dcache_inv>;
let SubtargetPredicate = isGFX7GFX8GFX9 in {
def S_DCACHE_INV_VOL : SM_Inval_Pseudo <"s_dcache_inv_vol", int_amdgcn_s_dcache_inv_vol>;
} // let SubtargetPredicate = isGFX7GFX8GFX9
let SubtargetPredicate = isGFX8Plus in {
let OtherPredicates = [HasScalarStores] in {
def S_DCACHE_WB : SM_Inval_Pseudo <"s_dcache_wb", int_amdgcn_s_dcache_wb>;
def S_DCACHE_WB_VOL : SM_Inval_Pseudo <"s_dcache_wb_vol", int_amdgcn_s_dcache_wb_vol>;
} // End OtherPredicates = [HasScalarStores]
def S_MEMREALTIME : SM_Time_Pseudo <"s_memrealtime", int_amdgcn_s_memrealtime>;
defm S_ATC_PROBE : SM_Pseudo_Probe <"s_atc_probe", SReg_64>;
let is_buffer = 1 in {
defm S_ATC_PROBE_BUFFER : SM_Pseudo_Probe <"s_atc_probe_buffer", SReg_128>;
}
} // SubtargetPredicate = isGFX8Plus
let SubtargetPredicate = isGFX10Plus in
def S_GL1_INV : SM_Inval_Pseudo<"s_gl1_inv">;
let SubtargetPredicate = HasGetWaveIdInst in
def S_GET_WAVEID_IN_WORKGROUP : SM_WaveId_Pseudo <"s_get_waveid_in_workgroup", int_amdgcn_s_get_waveid_in_workgroup>;
let SubtargetPredicate = HasScalarFlatScratchInsts, Uses = [FLAT_SCR] in {
defm S_SCRATCH_LOAD_DWORD : SM_Pseudo_Loads <"s_scratch_load_dword", SReg_64, SReg_32_XM0_XEXEC>;
defm S_SCRATCH_LOAD_DWORDX2 : SM_Pseudo_Loads <"s_scratch_load_dwordx2", SReg_64, SReg_64_XEXEC>;
defm S_SCRATCH_LOAD_DWORDX4 : SM_Pseudo_Loads <"s_scratch_load_dwordx4", SReg_64, SReg_128>;
defm S_SCRATCH_STORE_DWORD : SM_Pseudo_Stores <"s_scratch_store_dword", SReg_64, SReg_32_XM0_XEXEC>;
defm S_SCRATCH_STORE_DWORDX2 : SM_Pseudo_Stores <"s_scratch_store_dwordx2", SReg_64, SReg_64_XEXEC>;
defm S_SCRATCH_STORE_DWORDX4 : SM_Pseudo_Stores <"s_scratch_store_dwordx4", SReg_64, SReg_128>;
} // SubtargetPredicate = HasScalarFlatScratchInsts
let SubtargetPredicate = HasScalarAtomics in {
let is_buffer = 1 in {
defm S_BUFFER_ATOMIC_SWAP : SM_Pseudo_Atomics <"s_buffer_atomic_swap", SReg_128, SReg_32_XM0_XEXEC>;
defm S_BUFFER_ATOMIC_CMPSWAP : SM_Pseudo_Atomics <"s_buffer_atomic_cmpswap", SReg_128, SReg_64_XEXEC>;
defm S_BUFFER_ATOMIC_ADD : SM_Pseudo_Atomics <"s_buffer_atomic_add", SReg_128, SReg_32_XM0_XEXEC>;
defm S_BUFFER_ATOMIC_SUB : SM_Pseudo_Atomics <"s_buffer_atomic_sub", SReg_128, SReg_32_XM0_XEXEC>;
defm S_BUFFER_ATOMIC_SMIN : SM_Pseudo_Atomics <"s_buffer_atomic_smin", SReg_128, SReg_32_XM0_XEXEC>;
defm S_BUFFER_ATOMIC_UMIN : SM_Pseudo_Atomics <"s_buffer_atomic_umin", SReg_128, SReg_32_XM0_XEXEC>;
defm S_BUFFER_ATOMIC_SMAX : SM_Pseudo_Atomics <"s_buffer_atomic_smax", SReg_128, SReg_32_XM0_XEXEC>;
defm S_BUFFER_ATOMIC_UMAX : SM_Pseudo_Atomics <"s_buffer_atomic_umax", SReg_128, SReg_32_XM0_XEXEC>;
defm S_BUFFER_ATOMIC_AND : SM_Pseudo_Atomics <"s_buffer_atomic_and", SReg_128, SReg_32_XM0_XEXEC>;
defm S_BUFFER_ATOMIC_OR : SM_Pseudo_Atomics <"s_buffer_atomic_or", SReg_128, SReg_32_XM0_XEXEC>;
defm S_BUFFER_ATOMIC_XOR : SM_Pseudo_Atomics <"s_buffer_atomic_xor", SReg_128, SReg_32_XM0_XEXEC>;
defm S_BUFFER_ATOMIC_INC : SM_Pseudo_Atomics <"s_buffer_atomic_inc", SReg_128, SReg_32_XM0_XEXEC>;
defm S_BUFFER_ATOMIC_DEC : SM_Pseudo_Atomics <"s_buffer_atomic_dec", SReg_128, SReg_32_XM0_XEXEC>;
defm S_BUFFER_ATOMIC_SWAP_X2 : SM_Pseudo_Atomics <"s_buffer_atomic_swap_x2", SReg_128, SReg_64_XEXEC>;
defm S_BUFFER_ATOMIC_CMPSWAP_X2 : SM_Pseudo_Atomics <"s_buffer_atomic_cmpswap_x2", SReg_128, SReg_128>;
defm S_BUFFER_ATOMIC_ADD_X2 : SM_Pseudo_Atomics <"s_buffer_atomic_add_x2", SReg_128, SReg_64_XEXEC>;
defm S_BUFFER_ATOMIC_SUB_X2 : SM_Pseudo_Atomics <"s_buffer_atomic_sub_x2", SReg_128, SReg_64_XEXEC>;
defm S_BUFFER_ATOMIC_SMIN_X2 : SM_Pseudo_Atomics <"s_buffer_atomic_smin_x2", SReg_128, SReg_64_XEXEC>;
defm S_BUFFER_ATOMIC_UMIN_X2 : SM_Pseudo_Atomics <"s_buffer_atomic_umin_x2", SReg_128, SReg_64_XEXEC>;
defm S_BUFFER_ATOMIC_SMAX_X2 : SM_Pseudo_Atomics <"s_buffer_atomic_smax_x2", SReg_128, SReg_64_XEXEC>;
defm S_BUFFER_ATOMIC_UMAX_X2 : SM_Pseudo_Atomics <"s_buffer_atomic_umax_x2", SReg_128, SReg_64_XEXEC>;
defm S_BUFFER_ATOMIC_AND_X2 : SM_Pseudo_Atomics <"s_buffer_atomic_and_x2", SReg_128, SReg_64_XEXEC>;
defm S_BUFFER_ATOMIC_OR_X2 : SM_Pseudo_Atomics <"s_buffer_atomic_or_x2", SReg_128, SReg_64_XEXEC>;
defm S_BUFFER_ATOMIC_XOR_X2 : SM_Pseudo_Atomics <"s_buffer_atomic_xor_x2", SReg_128, SReg_64_XEXEC>;
defm S_BUFFER_ATOMIC_INC_X2 : SM_Pseudo_Atomics <"s_buffer_atomic_inc_x2", SReg_128, SReg_64_XEXEC>;
defm S_BUFFER_ATOMIC_DEC_X2 : SM_Pseudo_Atomics <"s_buffer_atomic_dec_x2", SReg_128, SReg_64_XEXEC>;
}
defm S_ATOMIC_SWAP : SM_Pseudo_Atomics <"s_atomic_swap", SReg_64, SReg_32_XM0_XEXEC>;
defm S_ATOMIC_CMPSWAP : SM_Pseudo_Atomics <"s_atomic_cmpswap", SReg_64, SReg_64_XEXEC>;
defm S_ATOMIC_ADD : SM_Pseudo_Atomics <"s_atomic_add", SReg_64, SReg_32_XM0_XEXEC>;
defm S_ATOMIC_SUB : SM_Pseudo_Atomics <"s_atomic_sub", SReg_64, SReg_32_XM0_XEXEC>;
defm S_ATOMIC_SMIN : SM_Pseudo_Atomics <"s_atomic_smin", SReg_64, SReg_32_XM0_XEXEC>;
defm S_ATOMIC_UMIN : SM_Pseudo_Atomics <"s_atomic_umin", SReg_64, SReg_32_XM0_XEXEC>;
defm S_ATOMIC_SMAX : SM_Pseudo_Atomics <"s_atomic_smax", SReg_64, SReg_32_XM0_XEXEC>;
defm S_ATOMIC_UMAX : SM_Pseudo_Atomics <"s_atomic_umax", SReg_64, SReg_32_XM0_XEXEC>;
defm S_ATOMIC_AND : SM_Pseudo_Atomics <"s_atomic_and", SReg_64, SReg_32_XM0_XEXEC>;
defm S_ATOMIC_OR : SM_Pseudo_Atomics <"s_atomic_or", SReg_64, SReg_32_XM0_XEXEC>;
defm S_ATOMIC_XOR : SM_Pseudo_Atomics <"s_atomic_xor", SReg_64, SReg_32_XM0_XEXEC>;
defm S_ATOMIC_INC : SM_Pseudo_Atomics <"s_atomic_inc", SReg_64, SReg_32_XM0_XEXEC>;
defm S_ATOMIC_DEC : SM_Pseudo_Atomics <"s_atomic_dec", SReg_64, SReg_32_XM0_XEXEC>;
defm S_ATOMIC_SWAP_X2 : SM_Pseudo_Atomics <"s_atomic_swap_x2", SReg_64, SReg_64_XEXEC>;
defm S_ATOMIC_CMPSWAP_X2 : SM_Pseudo_Atomics <"s_atomic_cmpswap_x2", SReg_64, SReg_128>;
defm S_ATOMIC_ADD_X2 : SM_Pseudo_Atomics <"s_atomic_add_x2", SReg_64, SReg_64_XEXEC>;
defm S_ATOMIC_SUB_X2 : SM_Pseudo_Atomics <"s_atomic_sub_x2", SReg_64, SReg_64_XEXEC>;
defm S_ATOMIC_SMIN_X2 : SM_Pseudo_Atomics <"s_atomic_smin_x2", SReg_64, SReg_64_XEXEC>;
defm S_ATOMIC_UMIN_X2 : SM_Pseudo_Atomics <"s_atomic_umin_x2", SReg_64, SReg_64_XEXEC>;
defm S_ATOMIC_SMAX_X2 : SM_Pseudo_Atomics <"s_atomic_smax_x2", SReg_64, SReg_64_XEXEC>;
defm S_ATOMIC_UMAX_X2 : SM_Pseudo_Atomics <"s_atomic_umax_x2", SReg_64, SReg_64_XEXEC>;
defm S_ATOMIC_AND_X2 : SM_Pseudo_Atomics <"s_atomic_and_x2", SReg_64, SReg_64_XEXEC>;
defm S_ATOMIC_OR_X2 : SM_Pseudo_Atomics <"s_atomic_or_x2", SReg_64, SReg_64_XEXEC>;
defm S_ATOMIC_XOR_X2 : SM_Pseudo_Atomics <"s_atomic_xor_x2", SReg_64, SReg_64_XEXEC>;
defm S_ATOMIC_INC_X2 : SM_Pseudo_Atomics <"s_atomic_inc_x2", SReg_64, SReg_64_XEXEC>;
defm S_ATOMIC_DEC_X2 : SM_Pseudo_Atomics <"s_atomic_dec_x2", SReg_64, SReg_64_XEXEC>;
} // let SubtargetPredicate = HasScalarAtomics
let SubtargetPredicate = HasScalarAtomics in {
defm S_DCACHE_DISCARD : SM_Pseudo_Discards <"s_dcache_discard">;
defm S_DCACHE_DISCARD_X2 : SM_Pseudo_Discards <"s_dcache_discard_x2">;
}
//===----------------------------------------------------------------------===//
// Targets
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// SI
//===----------------------------------------------------------------------===//
class SMRD_Real_si <bits<5> op, SM_Pseudo ps>
: SM_Real<ps>
, SIMCInstr<ps.PseudoInstr, SIEncodingFamily.SI>
, Enc32 {
let AssemblerPredicate = isGFX6GFX7;
let DecoderNamespace = "GFX6GFX7";
let Inst{7-0} = !if(ps.has_offset, offset{7-0}, ?);
let Inst{8} = imm;
let Inst{14-9} = !if(ps.has_sbase, sbase{6-1}, ?);
let Inst{21-15} = !if(ps.has_sdst, sdst{6-0}, ?);
let Inst{26-22} = op;
let Inst{31-27} = 0x18; //encoding
}
// FIXME: Assembler should reject trying to use glc on SMRD
// instructions on SI.
multiclass SM_Real_Loads_si<bits<5> op, string ps,
SM_Load_Pseudo immPs = !cast<SM_Load_Pseudo>(ps#_IMM),
SM_Load_Pseudo sgprPs = !cast<SM_Load_Pseudo>(ps#_SGPR)> {
def _IMM_si : SMRD_Real_si <op, immPs> {
let InOperandList = (ins immPs.BaseClass:$sbase, smrd_offset_8:$offset, GLC:$glc, DLC:$dlc);
}
// FIXME: The operand name $offset is inconsistent with $soff used
// in the pseudo
def _SGPR_si : SMRD_Real_si <op, sgprPs> {
let InOperandList = (ins sgprPs.BaseClass:$sbase, SReg_32:$offset, GLC:$glc, DLC:$dlc);
}
}
defm S_LOAD_DWORD : SM_Real_Loads_si <0x00, "S_LOAD_DWORD">;
defm S_LOAD_DWORDX2 : SM_Real_Loads_si <0x01, "S_LOAD_DWORDX2">;
defm S_LOAD_DWORDX4 : SM_Real_Loads_si <0x02, "S_LOAD_DWORDX4">;
defm S_LOAD_DWORDX8 : SM_Real_Loads_si <0x03, "S_LOAD_DWORDX8">;
defm S_LOAD_DWORDX16 : SM_Real_Loads_si <0x04, "S_LOAD_DWORDX16">;
defm S_BUFFER_LOAD_DWORD : SM_Real_Loads_si <0x08, "S_BUFFER_LOAD_DWORD">;
defm S_BUFFER_LOAD_DWORDX2 : SM_Real_Loads_si <0x09, "S_BUFFER_LOAD_DWORDX2">;
defm S_BUFFER_LOAD_DWORDX4 : SM_Real_Loads_si <0x0a, "S_BUFFER_LOAD_DWORDX4">;
defm S_BUFFER_LOAD_DWORDX8 : SM_Real_Loads_si <0x0b, "S_BUFFER_LOAD_DWORDX8">;
defm S_BUFFER_LOAD_DWORDX16 : SM_Real_Loads_si <0x0c, "S_BUFFER_LOAD_DWORDX16">;
def S_MEMTIME_si : SMRD_Real_si <0x1e, S_MEMTIME>;
def S_DCACHE_INV_si : SMRD_Real_si <0x1f, S_DCACHE_INV>;
//===----------------------------------------------------------------------===//
// VI
//===----------------------------------------------------------------------===//
class SMEM_Real_vi <bits<8> op, SM_Pseudo ps>
: SM_Real<ps>
, SIMCInstr<ps.PseudoInstr, SIEncodingFamily.VI>
, Enc64 {
bit glc;
let AssemblerPredicate = isGFX8GFX9;
let DecoderNamespace = "GFX8";
let Inst{5-0} = !if(ps.has_sbase, sbase{6-1}, ?);
let Inst{12-6} = !if(ps.has_sdst, sdst{6-0}, ?);
let Inst{16} = !if(ps.has_glc, glc, ?);
let Inst{17} = imm;
let Inst{25-18} = op;
let Inst{31-26} = 0x30; //encoding
// VI supports 20-bit unsigned offsets while GFX9+ supports 21-bit signed.
// Offset value is corrected accordingly when offset is encoded/decoded.
let Inst{52-32} = !if(ps.has_offset, offset{20-0}, ?);
}
multiclass SM_Real_Loads_vi<bits<8> op, string ps,
SM_Load_Pseudo immPs = !cast<SM_Load_Pseudo>(ps#_IMM),
SM_Load_Pseudo sgprPs = !cast<SM_Load_Pseudo>(ps#_SGPR)> {
def _IMM_vi : SMEM_Real_vi <op, immPs> {
let InOperandList = (ins immPs.BaseClass:$sbase, smem_offset:$offset, GLC:$glc, DLC:$dlc);
}
def _SGPR_vi : SMEM_Real_vi <op, sgprPs> {
let InOperandList = (ins sgprPs.BaseClass:$sbase, SReg_32:$offset, GLC:$glc, DLC:$dlc);
}
}
class SMEM_Real_Store_vi <bits<8> op, SM_Pseudo ps> : SMEM_Real_vi <op, ps> {
// encoding
bits<7> sdata;
let sdst = ?;
let Inst{12-6} = !if(ps.has_sdst, sdata{6-0}, ?);
}
multiclass SM_Real_Stores_vi<bits<8> op, string ps,
SM_Store_Pseudo immPs = !cast<SM_Store_Pseudo>(ps#_IMM),
SM_Store_Pseudo sgprPs = !cast<SM_Store_Pseudo>(ps#_SGPR)> {
// FIXME: The operand name $offset is inconsistent with $soff used
// in the pseudo
def _IMM_vi : SMEM_Real_Store_vi <op, immPs> {
let InOperandList = (ins immPs.SrcClass:$sdata, immPs.BaseClass:$sbase, smem_offset:$offset, GLC:$glc, DLC:$dlc);
}
def _SGPR_vi : SMEM_Real_Store_vi <op, sgprPs> {
let InOperandList = (ins sgprPs.SrcClass:$sdata, sgprPs.BaseClass:$sbase, SReg_32:$offset, GLC:$glc, DLC:$dlc);
}
}
multiclass SM_Real_Probe_vi<bits<8> op, string ps> {
def _IMM_vi : SMEM_Real_Store_vi <op, !cast<SM_Probe_Pseudo>(ps#_IMM)>;
def _SGPR_vi : SMEM_Real_Store_vi <op, !cast<SM_Probe_Pseudo>(ps#_SGPR)>;
}
defm S_LOAD_DWORD : SM_Real_Loads_vi <0x00, "S_LOAD_DWORD">;
defm S_LOAD_DWORDX2 : SM_Real_Loads_vi <0x01, "S_LOAD_DWORDX2">;
defm S_LOAD_DWORDX4 : SM_Real_Loads_vi <0x02, "S_LOAD_DWORDX4">;
defm S_LOAD_DWORDX8 : SM_Real_Loads_vi <0x03, "S_LOAD_DWORDX8">;
defm S_LOAD_DWORDX16 : SM_Real_Loads_vi <0x04, "S_LOAD_DWORDX16">;
defm S_BUFFER_LOAD_DWORD : SM_Real_Loads_vi <0x08, "S_BUFFER_LOAD_DWORD">;
defm S_BUFFER_LOAD_DWORDX2 : SM_Real_Loads_vi <0x09, "S_BUFFER_LOAD_DWORDX2">;
defm S_BUFFER_LOAD_DWORDX4 : SM_Real_Loads_vi <0x0a, "S_BUFFER_LOAD_DWORDX4">;
defm S_BUFFER_LOAD_DWORDX8 : SM_Real_Loads_vi <0x0b, "S_BUFFER_LOAD_DWORDX8">;
defm S_BUFFER_LOAD_DWORDX16 : SM_Real_Loads_vi <0x0c, "S_BUFFER_LOAD_DWORDX16">;
defm S_STORE_DWORD : SM_Real_Stores_vi <0x10, "S_STORE_DWORD">;
defm S_STORE_DWORDX2 : SM_Real_Stores_vi <0x11, "S_STORE_DWORDX2">;
defm S_STORE_DWORDX4 : SM_Real_Stores_vi <0x12, "S_STORE_DWORDX4">;
defm S_BUFFER_STORE_DWORD : SM_Real_Stores_vi <0x18, "S_BUFFER_STORE_DWORD">;
defm S_BUFFER_STORE_DWORDX2 : SM_Real_Stores_vi <0x19, "S_BUFFER_STORE_DWORDX2">;
defm S_BUFFER_STORE_DWORDX4 : SM_Real_Stores_vi <0x1a, "S_BUFFER_STORE_DWORDX4">;
// These instructions use same encoding
def S_DCACHE_INV_vi : SMEM_Real_vi <0x20, S_DCACHE_INV>;
def S_DCACHE_WB_vi : SMEM_Real_vi <0x21, S_DCACHE_WB>;
def S_DCACHE_INV_VOL_vi : SMEM_Real_vi <0x22, S_DCACHE_INV_VOL>;
def S_DCACHE_WB_VOL_vi : SMEM_Real_vi <0x23, S_DCACHE_WB_VOL>;
def S_MEMTIME_vi : SMEM_Real_vi <0x24, S_MEMTIME>;
def S_MEMREALTIME_vi : SMEM_Real_vi <0x25, S_MEMREALTIME>;
defm S_SCRATCH_LOAD_DWORD : SM_Real_Loads_vi <0x05, "S_SCRATCH_LOAD_DWORD">;
defm S_SCRATCH_LOAD_DWORDX2 : SM_Real_Loads_vi <0x06, "S_SCRATCH_LOAD_DWORDX2">;
defm S_SCRATCH_LOAD_DWORDX4 : SM_Real_Loads_vi <0x07, "S_SCRATCH_LOAD_DWORDX4">;
defm S_SCRATCH_STORE_DWORD : SM_Real_Stores_vi <0x15, "S_SCRATCH_STORE_DWORD">;
defm S_SCRATCH_STORE_DWORDX2 : SM_Real_Stores_vi <0x16, "S_SCRATCH_STORE_DWORDX2">;
defm S_SCRATCH_STORE_DWORDX4 : SM_Real_Stores_vi <0x17, "S_SCRATCH_STORE_DWORDX4">;
defm S_ATC_PROBE : SM_Real_Probe_vi <0x26, "S_ATC_PROBE">;
defm S_ATC_PROBE_BUFFER : SM_Real_Probe_vi <0x27, "S_ATC_PROBE_BUFFER">;
//===----------------------------------------------------------------------===//
// GFX9
//===----------------------------------------------------------------------===//
class SMEM_Atomic_Real_vi <bits<8> op, SM_Atomic_Pseudo ps>
: SMEM_Real_vi <op, ps> {
bits<7> sdata;
let Constraints = ps.Constraints;
let DisableEncoding = ps.DisableEncoding;
let glc = ps.glc;
let Inst{12-6} = !if(glc, sdst{6-0}, sdata{6-0});
}
multiclass SM_Real_Atomics_vi<bits<8> op, string ps> {
def _IMM_vi : SMEM_Atomic_Real_vi <op, !cast<SM_Atomic_Pseudo>(ps#_IMM)>;
def _SGPR_vi : SMEM_Atomic_Real_vi <op, !cast<SM_Atomic_Pseudo>(ps#_SGPR)>;
def _IMM_RTN_vi : SMEM_Atomic_Real_vi <op, !cast<SM_Atomic_Pseudo>(ps#_IMM_RTN)>;
def _SGPR_RTN_vi : SMEM_Atomic_Real_vi <op, !cast<SM_Atomic_Pseudo>(ps#_SGPR_RTN)>;
}
defm S_BUFFER_ATOMIC_SWAP : SM_Real_Atomics_vi <0x40, "S_BUFFER_ATOMIC_SWAP">;
defm S_BUFFER_ATOMIC_CMPSWAP : SM_Real_Atomics_vi <0x41, "S_BUFFER_ATOMIC_CMPSWAP">;
defm S_BUFFER_ATOMIC_ADD : SM_Real_Atomics_vi <0x42, "S_BUFFER_ATOMIC_ADD">;
defm S_BUFFER_ATOMIC_SUB : SM_Real_Atomics_vi <0x43, "S_BUFFER_ATOMIC_SUB">;
defm S_BUFFER_ATOMIC_SMIN : SM_Real_Atomics_vi <0x44, "S_BUFFER_ATOMIC_SMIN">;
defm S_BUFFER_ATOMIC_UMIN : SM_Real_Atomics_vi <0x45, "S_BUFFER_ATOMIC_UMIN">;
defm S_BUFFER_ATOMIC_SMAX : SM_Real_Atomics_vi <0x46, "S_BUFFER_ATOMIC_SMAX">;
defm S_BUFFER_ATOMIC_UMAX : SM_Real_Atomics_vi <0x47, "S_BUFFER_ATOMIC_UMAX">;
defm S_BUFFER_ATOMIC_AND : SM_Real_Atomics_vi <0x48, "S_BUFFER_ATOMIC_AND">;
defm S_BUFFER_ATOMIC_OR : SM_Real_Atomics_vi <0x49, "S_BUFFER_ATOMIC_OR">;
defm S_BUFFER_ATOMIC_XOR : SM_Real_Atomics_vi <0x4a, "S_BUFFER_ATOMIC_XOR">;
defm S_BUFFER_ATOMIC_INC : SM_Real_Atomics_vi <0x4b, "S_BUFFER_ATOMIC_INC">;
defm S_BUFFER_ATOMIC_DEC : SM_Real_Atomics_vi <0x4c, "S_BUFFER_ATOMIC_DEC">;
defm S_BUFFER_ATOMIC_SWAP_X2 : SM_Real_Atomics_vi <0x60, "S_BUFFER_ATOMIC_SWAP_X2">;
defm S_BUFFER_ATOMIC_CMPSWAP_X2 : SM_Real_Atomics_vi <0x61, "S_BUFFER_ATOMIC_CMPSWAP_X2">;
defm S_BUFFER_ATOMIC_ADD_X2 : SM_Real_Atomics_vi <0x62, "S_BUFFER_ATOMIC_ADD_X2">;
defm S_BUFFER_ATOMIC_SUB_X2 : SM_Real_Atomics_vi <0x63, "S_BUFFER_ATOMIC_SUB_X2">;
defm S_BUFFER_ATOMIC_SMIN_X2 : SM_Real_Atomics_vi <0x64, "S_BUFFER_ATOMIC_SMIN_X2">;
defm S_BUFFER_ATOMIC_UMIN_X2 : SM_Real_Atomics_vi <0x65, "S_BUFFER_ATOMIC_UMIN_X2">;
defm S_BUFFER_ATOMIC_SMAX_X2 : SM_Real_Atomics_vi <0x66, "S_BUFFER_ATOMIC_SMAX_X2">;
defm S_BUFFER_ATOMIC_UMAX_X2 : SM_Real_Atomics_vi <0x67, "S_BUFFER_ATOMIC_UMAX_X2">;
defm S_BUFFER_ATOMIC_AND_X2 : SM_Real_Atomics_vi <0x68, "S_BUFFER_ATOMIC_AND_X2">;
defm S_BUFFER_ATOMIC_OR_X2 : SM_Real_Atomics_vi <0x69, "S_BUFFER_ATOMIC_OR_X2">;
defm S_BUFFER_ATOMIC_XOR_X2 : SM_Real_Atomics_vi <0x6a, "S_BUFFER_ATOMIC_XOR_X2">;
defm S_BUFFER_ATOMIC_INC_X2 : SM_Real_Atomics_vi <0x6b, "S_BUFFER_ATOMIC_INC_X2">;
defm S_BUFFER_ATOMIC_DEC_X2 : SM_Real_Atomics_vi <0x6c, "S_BUFFER_ATOMIC_DEC_X2">;
defm S_ATOMIC_SWAP : SM_Real_Atomics_vi <0x80, "S_ATOMIC_SWAP">;
defm S_ATOMIC_CMPSWAP : SM_Real_Atomics_vi <0x81, "S_ATOMIC_CMPSWAP">;
defm S_ATOMIC_ADD : SM_Real_Atomics_vi <0x82, "S_ATOMIC_ADD">;
defm S_ATOMIC_SUB : SM_Real_Atomics_vi <0x83, "S_ATOMIC_SUB">;
defm S_ATOMIC_SMIN : SM_Real_Atomics_vi <0x84, "S_ATOMIC_SMIN">;
defm S_ATOMIC_UMIN : SM_Real_Atomics_vi <0x85, "S_ATOMIC_UMIN">;
defm S_ATOMIC_SMAX : SM_Real_Atomics_vi <0x86, "S_ATOMIC_SMAX">;
defm S_ATOMIC_UMAX : SM_Real_Atomics_vi <0x87, "S_ATOMIC_UMAX">;
defm S_ATOMIC_AND : SM_Real_Atomics_vi <0x88, "S_ATOMIC_AND">;
defm S_ATOMIC_OR : SM_Real_Atomics_vi <0x89, "S_ATOMIC_OR">;
defm S_ATOMIC_XOR : SM_Real_Atomics_vi <0x8a, "S_ATOMIC_XOR">;
defm S_ATOMIC_INC : SM_Real_Atomics_vi <0x8b, "S_ATOMIC_INC">;
defm S_ATOMIC_DEC : SM_Real_Atomics_vi <0x8c, "S_ATOMIC_DEC">;
defm S_ATOMIC_SWAP_X2 : SM_Real_Atomics_vi <0xa0, "S_ATOMIC_SWAP_X2">;
defm S_ATOMIC_CMPSWAP_X2 : SM_Real_Atomics_vi <0xa1, "S_ATOMIC_CMPSWAP_X2">;
defm S_ATOMIC_ADD_X2 : SM_Real_Atomics_vi <0xa2, "S_ATOMIC_ADD_X2">;
defm S_ATOMIC_SUB_X2 : SM_Real_Atomics_vi <0xa3, "S_ATOMIC_SUB_X2">;
defm S_ATOMIC_SMIN_X2 : SM_Real_Atomics_vi <0xa4, "S_ATOMIC_SMIN_X2">;
defm S_ATOMIC_UMIN_X2 : SM_Real_Atomics_vi <0xa5, "S_ATOMIC_UMIN_X2">;
defm S_ATOMIC_SMAX_X2 : SM_Real_Atomics_vi <0xa6, "S_ATOMIC_SMAX_X2">;
defm S_ATOMIC_UMAX_X2 : SM_Real_Atomics_vi <0xa7, "S_ATOMIC_UMAX_X2">;
defm S_ATOMIC_AND_X2 : SM_Real_Atomics_vi <0xa8, "S_ATOMIC_AND_X2">;
defm S_ATOMIC_OR_X2 : SM_Real_Atomics_vi <0xa9, "S_ATOMIC_OR_X2">;
defm S_ATOMIC_XOR_X2 : SM_Real_Atomics_vi <0xaa, "S_ATOMIC_XOR_X2">;
defm S_ATOMIC_INC_X2 : SM_Real_Atomics_vi <0xab, "S_ATOMIC_INC_X2">;
defm S_ATOMIC_DEC_X2 : SM_Real_Atomics_vi <0xac, "S_ATOMIC_DEC_X2">;
multiclass SM_Real_Discard_vi<bits<8> op, string ps> {
def _IMM_vi : SMEM_Real_vi <op, !cast<SM_Discard_Pseudo>(ps#_IMM)>;
def _SGPR_vi : SMEM_Real_vi <op, !cast<SM_Discard_Pseudo>(ps#_SGPR)>;
}
defm S_DCACHE_DISCARD : SM_Real_Discard_vi <0x28, "S_DCACHE_DISCARD">;
defm S_DCACHE_DISCARD_X2 : SM_Real_Discard_vi <0x29, "S_DCACHE_DISCARD_X2">;
//===----------------------------------------------------------------------===//
// CI
//===----------------------------------------------------------------------===//
def smrd_literal_offset : NamedOperandU32<"SMRDLiteralOffset",
NamedMatchClass<"SMRDLiteralOffset">> {
let OperandType = "OPERAND_IMMEDIATE";
}
class SMRD_Real_Load_IMM_ci <bits<5> op, SM_Load_Pseudo ps> :
SM_Real<ps>,
Enc64 {
let AssemblerPredicate = isGFX7Only;
let DecoderNamespace = "GFX7";
let InOperandList = (ins ps.BaseClass:$sbase, smrd_literal_offset:$offset, GLC:$glc, DLC:$dlc);
let LGKM_CNT = ps.LGKM_CNT;
let mayLoad = ps.mayLoad;
let mayStore = ps.mayStore;
let hasSideEffects = ps.hasSideEffects;
let SchedRW = ps.SchedRW;
let Inst{7-0} = 0xff;
let Inst{8} = 0;
let Inst{14-9} = sbase{6-1};
let Inst{21-15} = sdst{6-0};
let Inst{26-22} = op;
let Inst{31-27} = 0x18; //encoding
let Inst{63-32} = offset{31-0};
}
def S_LOAD_DWORD_IMM_ci : SMRD_Real_Load_IMM_ci <0x00, S_LOAD_DWORD_IMM>;
def S_LOAD_DWORDX2_IMM_ci : SMRD_Real_Load_IMM_ci <0x01, S_LOAD_DWORDX2_IMM>;
def S_LOAD_DWORDX4_IMM_ci : SMRD_Real_Load_IMM_ci <0x02, S_LOAD_DWORDX4_IMM>;
def S_LOAD_DWORDX8_IMM_ci : SMRD_Real_Load_IMM_ci <0x03, S_LOAD_DWORDX8_IMM>;
def S_LOAD_DWORDX16_IMM_ci : SMRD_Real_Load_IMM_ci <0x04, S_LOAD_DWORDX16_IMM>;
def S_BUFFER_LOAD_DWORD_IMM_ci : SMRD_Real_Load_IMM_ci <0x08, S_BUFFER_LOAD_DWORD_IMM>;
def S_BUFFER_LOAD_DWORDX2_IMM_ci : SMRD_Real_Load_IMM_ci <0x09, S_BUFFER_LOAD_DWORDX2_IMM>;
def S_BUFFER_LOAD_DWORDX4_IMM_ci : SMRD_Real_Load_IMM_ci <0x0a, S_BUFFER_LOAD_DWORDX4_IMM>;
def S_BUFFER_LOAD_DWORDX8_IMM_ci : SMRD_Real_Load_IMM_ci <0x0b, S_BUFFER_LOAD_DWORDX8_IMM>;
def S_BUFFER_LOAD_DWORDX16_IMM_ci : SMRD_Real_Load_IMM_ci <0x0c, S_BUFFER_LOAD_DWORDX16_IMM>;
class SMRD_Real_ci <bits<5> op, SM_Pseudo ps>
: SM_Real<ps>
, SIMCInstr<ps.PseudoInstr, SIEncodingFamily.SI>
, Enc32 {
let AssemblerPredicate = isGFX7Only;
let DecoderNamespace = "GFX7";
let Inst{7-0} = !if(ps.has_offset, offset{7-0}, ?);
let Inst{8} = imm;
let Inst{14-9} = !if(ps.has_sbase, sbase{6-1}, ?);
let Inst{21-15} = !if(ps.has_sdst, sdst{6-0}, ?);
let Inst{26-22} = op;
let Inst{31-27} = 0x18; //encoding
}
def S_DCACHE_INV_VOL_ci : SMRD_Real_ci <0x1d, S_DCACHE_INV_VOL>;
//===----------------------------------------------------------------------===//
// Scalar Memory Patterns
//===----------------------------------------------------------------------===//
def smrd_load : PatFrag <(ops node:$ptr), (load node:$ptr), [{ return isUniformLoad(N);}]> {
let GISelPredicateCode = [{
if (!MI.hasOneMemOperand())
return false;
if (!isInstrUniform(MI))
return false;
// FIXME: We should probably be caching this.
SmallVector<GEPInfo, 4> AddrInfo;
getAddrModeInfo(MI, MRI, AddrInfo);
if (hasVgprParts(AddrInfo))
return false;
return true;
}];
}
def SMRDImm : ComplexPattern<i64, 2, "SelectSMRDImm">;
def SMRDImm32 : ComplexPattern<i64, 2, "SelectSMRDImm32">;
def SMRDSgpr : ComplexPattern<i64, 2, "SelectSMRDSgpr">;
def SMRDBufferImm : ComplexPattern<i32, 1, "SelectSMRDBufferImm">;
def SMRDBufferImm32 : ComplexPattern<i32, 1, "SelectSMRDBufferImm32">;
multiclass SMRD_Pattern <string Instr, ValueType vt> {
// 1. IMM offset
def : GCNPat <
(smrd_load (SMRDImm i64:$sbase, i32:$offset)),
(vt (!cast<SM_Pseudo>(Instr#"_IMM") $sbase, $offset, 0, 0))
>;
// 2. 32-bit IMM offset on CI
def : GCNPat <
(smrd_load (SMRDImm32 i64:$sbase, i32:$offset)),
(vt (!cast<InstSI>(Instr#"_IMM_ci") $sbase, $offset, 0, 0))> {
let OtherPredicates = [isGFX7Only];
}
// 3. SGPR offset
def : GCNPat <
(smrd_load (SMRDSgpr i64:$sbase, i32:$offset)),
(vt (!cast<SM_Pseudo>(Instr#"_SGPR") $sbase, $offset, 0, 0))
>;
// 4. No offset
def : GCNPat <
(vt (smrd_load (i64 SReg_64:$sbase))),
(vt (!cast<SM_Pseudo>(Instr#"_IMM") i64:$sbase, 0, 0, 0))
>;
}
multiclass SMLoad_Pattern <string Instr, ValueType vt> {
// 1. Offset as an immediate
def : GCNPat <
(SIsbuffer_load v4i32:$sbase, (SMRDBufferImm i32:$offset), timm:$cachepolicy),
(vt (!cast<SM_Pseudo>(Instr#"_IMM") SReg_128:$sbase, i32imm:$offset, (extract_glc $cachepolicy),
(extract_dlc $cachepolicy)))> {
let AddedComplexity = 2;
}
// 2. 32-bit IMM offset on CI
def : GCNPat <
(vt (SIsbuffer_load v4i32:$sbase, (SMRDBufferImm32 i32:$offset), timm:$cachepolicy)),
(!cast<InstSI>(Instr#"_IMM_ci") SReg_128:$sbase, smrd_literal_offset:$offset,
(extract_glc $cachepolicy), (extract_dlc $cachepolicy))> {
let OtherPredicates = [isGFX7Only];
let AddedComplexity = 1;
}
// 3. Offset loaded in an 32bit SGPR
def : GCNPat <
(SIsbuffer_load v4i32:$sbase, i32:$offset, timm:$cachepolicy),
(vt (!cast<SM_Pseudo>(Instr#"_SGPR") SReg_128:$sbase, SReg_32:$offset, (extract_glc $cachepolicy),
(extract_dlc $cachepolicy)))
>;
}
// Global and constant loads can be selected to either MUBUF or SMRD
// instructions, but SMRD instructions are faster so we want the instruction
// selector to prefer those.
let AddedComplexity = 100 in {
foreach vt = Reg32Types.types in {
defm : SMRD_Pattern <"S_LOAD_DWORD", vt>;
}
foreach vt = SReg_64.RegTypes in {
defm : SMRD_Pattern <"S_LOAD_DWORDX2", vt>;
}
foreach vt = SReg_128.RegTypes in {
defm : SMRD_Pattern <"S_LOAD_DWORDX4", vt>;
}
foreach vt = SReg_256.RegTypes in {
defm : SMRD_Pattern <"S_LOAD_DWORDX8", vt>;
}
foreach vt = SReg_512.RegTypes in {
defm : SMRD_Pattern <"S_LOAD_DWORDX16", vt>;
}
defm : SMLoad_Pattern <"S_BUFFER_LOAD_DWORD", i32>;
defm : SMLoad_Pattern <"S_BUFFER_LOAD_DWORDX2", v2i32>;
defm : SMLoad_Pattern <"S_BUFFER_LOAD_DWORDX4", v4i32>;
defm : SMLoad_Pattern <"S_BUFFER_LOAD_DWORDX8", v8i32>;
defm : SMLoad_Pattern <"S_BUFFER_LOAD_DWORDX16", v16i32>;
defm : SMLoad_Pattern <"S_BUFFER_LOAD_DWORD", f32>;
defm : SMLoad_Pattern <"S_BUFFER_LOAD_DWORDX2", v2f32>;
defm : SMLoad_Pattern <"S_BUFFER_LOAD_DWORDX4", v4f32>;
defm : SMLoad_Pattern <"S_BUFFER_LOAD_DWORDX8", v8f32>;
defm : SMLoad_Pattern <"S_BUFFER_LOAD_DWORDX16", v16f32>;
} // End let AddedComplexity = 100
let OtherPredicates = [HasSMemTimeInst] in {
def : GCNPat <
(i64 (readcyclecounter)),
(S_MEMTIME)
>;
} // let OtherPredicates = [HasSMemTimeInst]
let OtherPredicates = [HasNoSMemTimeInst] in {
def : GCNPat <
(i64 (readcyclecounter)),
(REG_SEQUENCE SReg_64,
(S_GETREG_B32 getHwRegImm<HWREG.SHADER_CYCLES, 0, -12>.ret), sub0,
(S_MOV_B32 (i32 0)), sub1)
>;
} // let OtherPredicates = [HasNoSMemTimeInst]
//===----------------------------------------------------------------------===//
// GFX10.
//===----------------------------------------------------------------------===//
class SMEM_Real_gfx10<bits<8> op, SM_Pseudo ps> :
SM_Real<ps>, SIMCInstr<ps.PseudoInstr, SIEncodingFamily.GFX10>, Enc64 {
bit glc;
bit dlc;
let AssemblerPredicate = isGFX10Plus;
let DecoderNamespace = "GFX10";
let Inst{5-0} = !if(ps.has_sbase, sbase{6-1}, ?);
let Inst{12-6} = !if(ps.has_sdst, sdst{6-0}, ?);
let Inst{14} = !if(ps.has_dlc, dlc, ?);
let Inst{16} = !if(ps.has_glc, glc, ?);
let Inst{25-18} = op;
let Inst{31-26} = 0x3d;
let Inst{52-32} = !if(ps.offset_is_imm, !if(ps.has_offset, offset{20-0}, ?), ?);
let Inst{63-57} = !if(ps.offset_is_imm, !cast<int>(SGPR_NULL.HWEncoding),
!if(ps.has_offset, offset{6-0}, ?));
}
multiclass SM_Real_Loads_gfx10<bits<8> op, string ps,
SM_Load_Pseudo immPs = !cast<SM_Load_Pseudo>(ps#_IMM),
SM_Load_Pseudo sgprPs = !cast<SM_Load_Pseudo>(ps#_SGPR)> {
def _IMM_gfx10 : SMEM_Real_gfx10<op, immPs> {
let InOperandList = (ins immPs.BaseClass:$sbase, smem_offset:$offset, GLC:$glc, DLC:$dlc);
}
def _SGPR_gfx10 : SMEM_Real_gfx10<op, sgprPs> {
let InOperandList = (ins sgprPs.BaseClass:$sbase, SReg_32:$offset, GLC:$glc, DLC:$dlc);
}
}
class SMEM_Real_Store_gfx10<bits<8> op, SM_Pseudo ps> : SMEM_Real_gfx10<op, ps> {
bits<7> sdata;
let sdst = ?;
let Inst{12-6} = !if(ps.has_sdst, sdata{6-0}, ?);
}
multiclass SM_Real_Stores_gfx10<bits<8> op, string ps,
SM_Store_Pseudo immPs = !cast<SM_Store_Pseudo>(ps#_IMM),
SM_Store_Pseudo sgprPs = !cast<SM_Store_Pseudo>(ps#_SGPR)> {
// FIXME: The operand name $offset is inconsistent with $soff used
// in the pseudo
def _IMM_gfx10 : SMEM_Real_Store_gfx10 <op, immPs> {
let InOperandList = (ins immPs.SrcClass:$sdata, immPs.BaseClass:$sbase, smem_offset:$offset, GLC:$glc, DLC:$dlc);
}
def _SGPR_gfx10 : SMEM_Real_Store_gfx10 <op, sgprPs> {
let InOperandList = (ins sgprPs.SrcClass:$sdata, sgprPs.BaseClass:$sbase, SReg_32:$offset, GLC:$glc, DLC:$dlc);
}
}
defm S_LOAD_DWORD : SM_Real_Loads_gfx10<0x000, "S_LOAD_DWORD">;
defm S_LOAD_DWORDX2 : SM_Real_Loads_gfx10<0x001, "S_LOAD_DWORDX2">;
defm S_LOAD_DWORDX4 : SM_Real_Loads_gfx10<0x002, "S_LOAD_DWORDX4">;
defm S_LOAD_DWORDX8 : SM_Real_Loads_gfx10<0x003, "S_LOAD_DWORDX8">;
defm S_LOAD_DWORDX16 : SM_Real_Loads_gfx10<0x004, "S_LOAD_DWORDX16">;
let SubtargetPredicate = HasScalarFlatScratchInsts in {
defm S_SCRATCH_LOAD_DWORD : SM_Real_Loads_gfx10<0x005, "S_SCRATCH_LOAD_DWORD">;
defm S_SCRATCH_LOAD_DWORDX2 : SM_Real_Loads_gfx10<0x006, "S_SCRATCH_LOAD_DWORDX2">;
defm S_SCRATCH_LOAD_DWORDX4 : SM_Real_Loads_gfx10<0x007, "S_SCRATCH_LOAD_DWORDX4">;
} // End SubtargetPredicate = HasScalarFlatScratchInsts
defm S_BUFFER_LOAD_DWORD : SM_Real_Loads_gfx10<0x008, "S_BUFFER_LOAD_DWORD">;
defm S_BUFFER_LOAD_DWORDX2 : SM_Real_Loads_gfx10<0x009, "S_BUFFER_LOAD_DWORDX2">;
defm S_BUFFER_LOAD_DWORDX4 : SM_Real_Loads_gfx10<0x00a, "S_BUFFER_LOAD_DWORDX4">;
defm S_BUFFER_LOAD_DWORDX8 : SM_Real_Loads_gfx10<0x00b, "S_BUFFER_LOAD_DWORDX8">;
defm S_BUFFER_LOAD_DWORDX16 : SM_Real_Loads_gfx10<0x00c, "S_BUFFER_LOAD_DWORDX16">;
let SubtargetPredicate = HasScalarStores in {
defm S_STORE_DWORD : SM_Real_Stores_gfx10<0x010, "S_STORE_DWORD">;
defm S_STORE_DWORDX2 : SM_Real_Stores_gfx10<0x011, "S_STORE_DWORDX2">;
defm S_STORE_DWORDX4 : SM_Real_Stores_gfx10<0x012, "S_STORE_DWORDX4">;
let OtherPredicates = [HasScalarFlatScratchInsts] in {
defm S_SCRATCH_STORE_DWORD : SM_Real_Stores_gfx10<0x015, "S_SCRATCH_STORE_DWORD">;
defm S_SCRATCH_STORE_DWORDX2 : SM_Real_Stores_gfx10<0x016, "S_SCRATCH_STORE_DWORDX2">;
defm S_SCRATCH_STORE_DWORDX4 : SM_Real_Stores_gfx10<0x017, "S_SCRATCH_STORE_DWORDX4">;
} // End OtherPredicates = [HasScalarFlatScratchInsts]
defm S_BUFFER_STORE_DWORD : SM_Real_Stores_gfx10<0x018, "S_BUFFER_STORE_DWORD">;
defm S_BUFFER_STORE_DWORDX2 : SM_Real_Stores_gfx10<0x019, "S_BUFFER_STORE_DWORDX2">;
defm S_BUFFER_STORE_DWORDX4 : SM_Real_Stores_gfx10<0x01a, "S_BUFFER_STORE_DWORDX4">;
} // End SubtargetPredicate = HasScalarStores
def S_MEMREALTIME_gfx10 : SMEM_Real_gfx10<0x025, S_MEMREALTIME>;
def S_MEMTIME_gfx10 : SMEM_Real_gfx10<0x024, S_MEMTIME>;
def S_GL1_INV_gfx10 : SMEM_Real_gfx10<0x01f, S_GL1_INV>;
def S_GET_WAVEID_IN_WORKGROUP_gfx10 : SMEM_Real_gfx10<0x02a, S_GET_WAVEID_IN_WORKGROUP>;
def S_DCACHE_INV_gfx10 : SMEM_Real_gfx10<0x020, S_DCACHE_INV>;
let SubtargetPredicate = HasScalarStores in {
def S_DCACHE_WB_gfx10 : SMEM_Real_gfx10<0x021, S_DCACHE_WB>;
} // End SubtargetPredicate = HasScalarStores
multiclass SM_Real_Probe_gfx10<bits<8> op, string ps> {
def _IMM_gfx10 : SMEM_Real_Store_gfx10 <op, !cast<SM_Pseudo>(ps#_IMM)>;
def _SGPR_gfx10 : SMEM_Real_Store_gfx10 <op, !cast<SM_Pseudo>(ps#_SGPR)>;
}
defm S_ATC_PROBE : SM_Real_Probe_gfx10 <0x26, "S_ATC_PROBE">;
defm S_ATC_PROBE_BUFFER : SM_Real_Probe_gfx10 <0x27, "S_ATC_PROBE_BUFFER">;
class SMEM_Atomic_Real_gfx10 <bits<8> op, SM_Atomic_Pseudo ps>
: SMEM_Real_gfx10 <op, ps> {
bits<7> sdata;
bit dlc;
let Constraints = ps.Constraints;
let DisableEncoding = ps.DisableEncoding;
let glc = ps.glc;
let Inst{14} = !if(ps.has_dlc, dlc, 0);
let Inst{12-6} = !if(glc, sdst{6-0}, sdata{6-0});
}
multiclass SM_Real_Atomics_gfx10<bits<8> op, string ps> {
def _IMM_gfx10 : SMEM_Atomic_Real_gfx10 <op, !cast<SM_Atomic_Pseudo>(ps#_IMM)>;
def _SGPR_gfx10 : SMEM_Atomic_Real_gfx10 <op, !cast<SM_Atomic_Pseudo>(ps#_SGPR)>;
def _IMM_RTN_gfx10 : SMEM_Atomic_Real_gfx10 <op, !cast<SM_Atomic_Pseudo>(ps#_IMM_RTN)>;
def _SGPR_RTN_gfx10 : SMEM_Atomic_Real_gfx10 <op, !cast<SM_Atomic_Pseudo>(ps#_SGPR_RTN)>;
}
let SubtargetPredicate = HasScalarAtomics in {
defm S_BUFFER_ATOMIC_SWAP : SM_Real_Atomics_gfx10 <0x40, "S_BUFFER_ATOMIC_SWAP">;
defm S_BUFFER_ATOMIC_CMPSWAP : SM_Real_Atomics_gfx10 <0x41, "S_BUFFER_ATOMIC_CMPSWAP">;
defm S_BUFFER_ATOMIC_ADD : SM_Real_Atomics_gfx10 <0x42, "S_BUFFER_ATOMIC_ADD">;
defm S_BUFFER_ATOMIC_SUB : SM_Real_Atomics_gfx10 <0x43, "S_BUFFER_ATOMIC_SUB">;
defm S_BUFFER_ATOMIC_SMIN : SM_Real_Atomics_gfx10 <0x44, "S_BUFFER_ATOMIC_SMIN">;
defm S_BUFFER_ATOMIC_UMIN : SM_Real_Atomics_gfx10 <0x45, "S_BUFFER_ATOMIC_UMIN">;
defm S_BUFFER_ATOMIC_SMAX : SM_Real_Atomics_gfx10 <0x46, "S_BUFFER_ATOMIC_SMAX">;
defm S_BUFFER_ATOMIC_UMAX : SM_Real_Atomics_gfx10 <0x47, "S_BUFFER_ATOMIC_UMAX">;
defm S_BUFFER_ATOMIC_AND : SM_Real_Atomics_gfx10 <0x48, "S_BUFFER_ATOMIC_AND">;
defm S_BUFFER_ATOMIC_OR : SM_Real_Atomics_gfx10 <0x49, "S_BUFFER_ATOMIC_OR">;
defm S_BUFFER_ATOMIC_XOR : SM_Real_Atomics_gfx10 <0x4a, "S_BUFFER_ATOMIC_XOR">;
defm S_BUFFER_ATOMIC_INC : SM_Real_Atomics_gfx10 <0x4b, "S_BUFFER_ATOMIC_INC">;
defm S_BUFFER_ATOMIC_DEC : SM_Real_Atomics_gfx10 <0x4c, "S_BUFFER_ATOMIC_DEC">;
defm S_BUFFER_ATOMIC_SWAP_X2 : SM_Real_Atomics_gfx10 <0x60, "S_BUFFER_ATOMIC_SWAP_X2">;
defm S_BUFFER_ATOMIC_CMPSWAP_X2 : SM_Real_Atomics_gfx10 <0x61, "S_BUFFER_ATOMIC_CMPSWAP_X2">;
defm S_BUFFER_ATOMIC_ADD_X2 : SM_Real_Atomics_gfx10 <0x62, "S_BUFFER_ATOMIC_ADD_X2">;
defm S_BUFFER_ATOMIC_SUB_X2 : SM_Real_Atomics_gfx10 <0x63, "S_BUFFER_ATOMIC_SUB_X2">;
defm S_BUFFER_ATOMIC_SMIN_X2 : SM_Real_Atomics_gfx10 <0x64, "S_BUFFER_ATOMIC_SMIN_X2">;
defm S_BUFFER_ATOMIC_UMIN_X2 : SM_Real_Atomics_gfx10 <0x65, "S_BUFFER_ATOMIC_UMIN_X2">;
defm S_BUFFER_ATOMIC_SMAX_X2 : SM_Real_Atomics_gfx10 <0x66, "S_BUFFER_ATOMIC_SMAX_X2">;
defm S_BUFFER_ATOMIC_UMAX_X2 : SM_Real_Atomics_gfx10 <0x67, "S_BUFFER_ATOMIC_UMAX_X2">;
defm S_BUFFER_ATOMIC_AND_X2 : SM_Real_Atomics_gfx10 <0x68, "S_BUFFER_ATOMIC_AND_X2">;
defm S_BUFFER_ATOMIC_OR_X2 : SM_Real_Atomics_gfx10 <0x69, "S_BUFFER_ATOMIC_OR_X2">;
defm S_BUFFER_ATOMIC_XOR_X2 : SM_Real_Atomics_gfx10 <0x6a, "S_BUFFER_ATOMIC_XOR_X2">;
defm S_BUFFER_ATOMIC_INC_X2 : SM_Real_Atomics_gfx10 <0x6b, "S_BUFFER_ATOMIC_INC_X2">;
defm S_BUFFER_ATOMIC_DEC_X2 : SM_Real_Atomics_gfx10 <0x6c, "S_BUFFER_ATOMIC_DEC_X2">;
defm S_ATOMIC_SWAP : SM_Real_Atomics_gfx10 <0x80, "S_ATOMIC_SWAP">;
defm S_ATOMIC_CMPSWAP : SM_Real_Atomics_gfx10 <0x81, "S_ATOMIC_CMPSWAP">;
defm S_ATOMIC_ADD : SM_Real_Atomics_gfx10 <0x82, "S_ATOMIC_ADD">;
defm S_ATOMIC_SUB : SM_Real_Atomics_gfx10 <0x83, "S_ATOMIC_SUB">;
defm S_ATOMIC_SMIN : SM_Real_Atomics_gfx10 <0x84, "S_ATOMIC_SMIN">;
defm S_ATOMIC_UMIN : SM_Real_Atomics_gfx10 <0x85, "S_ATOMIC_UMIN">;
defm S_ATOMIC_SMAX : SM_Real_Atomics_gfx10 <0x86, "S_ATOMIC_SMAX">;
defm S_ATOMIC_UMAX : SM_Real_Atomics_gfx10 <0x87, "S_ATOMIC_UMAX">;
defm S_ATOMIC_AND : SM_Real_Atomics_gfx10 <0x88, "S_ATOMIC_AND">;
defm S_ATOMIC_OR : SM_Real_Atomics_gfx10 <0x89, "S_ATOMIC_OR">;
defm S_ATOMIC_XOR : SM_Real_Atomics_gfx10 <0x8a, "S_ATOMIC_XOR">;
defm S_ATOMIC_INC : SM_Real_Atomics_gfx10 <0x8b, "S_ATOMIC_INC">;
defm S_ATOMIC_DEC : SM_Real_Atomics_gfx10 <0x8c, "S_ATOMIC_DEC">;
defm S_ATOMIC_SWAP_X2 : SM_Real_Atomics_gfx10 <0xa0, "S_ATOMIC_SWAP_X2">;
defm S_ATOMIC_CMPSWAP_X2 : SM_Real_Atomics_gfx10 <0xa1, "S_ATOMIC_CMPSWAP_X2">;
defm S_ATOMIC_ADD_X2 : SM_Real_Atomics_gfx10 <0xa2, "S_ATOMIC_ADD_X2">;
defm S_ATOMIC_SUB_X2 : SM_Real_Atomics_gfx10 <0xa3, "S_ATOMIC_SUB_X2">;
defm S_ATOMIC_SMIN_X2 : SM_Real_Atomics_gfx10 <0xa4, "S_ATOMIC_SMIN_X2">;
defm S_ATOMIC_UMIN_X2 : SM_Real_Atomics_gfx10 <0xa5, "S_ATOMIC_UMIN_X2">;
defm S_ATOMIC_SMAX_X2 : SM_Real_Atomics_gfx10 <0xa6, "S_ATOMIC_SMAX_X2">;
defm S_ATOMIC_UMAX_X2 : SM_Real_Atomics_gfx10 <0xa7, "S_ATOMIC_UMAX_X2">;
defm S_ATOMIC_AND_X2 : SM_Real_Atomics_gfx10 <0xa8, "S_ATOMIC_AND_X2">;
defm S_ATOMIC_OR_X2 : SM_Real_Atomics_gfx10 <0xa9, "S_ATOMIC_OR_X2">;
defm S_ATOMIC_XOR_X2 : SM_Real_Atomics_gfx10 <0xaa, "S_ATOMIC_XOR_X2">;
defm S_ATOMIC_INC_X2 : SM_Real_Atomics_gfx10 <0xab, "S_ATOMIC_INC_X2">;
defm S_ATOMIC_DEC_X2 : SM_Real_Atomics_gfx10 <0xac, "S_ATOMIC_DEC_X2">;
multiclass SM_Real_Discard_gfx10<bits<8> op, string ps> {
def _IMM_gfx10 : SMEM_Real_gfx10 <op, !cast<SM_Pseudo>(ps#_IMM)>;
def _SGPR_gfx10 : SMEM_Real_gfx10 <op, !cast<SM_Pseudo>(ps#_SGPR)>;
}
defm S_DCACHE_DISCARD : SM_Real_Discard_gfx10 <0x28, "S_DCACHE_DISCARD">;
defm S_DCACHE_DISCARD_X2 : SM_Real_Discard_gfx10 <0x29, "S_DCACHE_DISCARD_X2">;
} // End SubtargetPredicate = HasScalarAtomics
def SMInfoTable : GenericTable {
let FilterClass = "SM_Real";
let CppTypeName = "SMInfo";
let Fields = ["Opcode", "is_buffer"];
let PrimaryKey = ["Opcode"];
let PrimaryKeyName = "getSMEMOpcodeHelper";
}