blob: 8d69f7a515333db04f7a791c0d3ec1c382df5812 [file] [log] [blame]
//===- NVPTXIntrinsics.td - PTX Intrinsics Instructions -------*- tblgen -*-==//
//
// 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 immFloat0 : PatLeaf<(fpimm), [{
float f = (float)N->getValueAPF().convertToFloat();
return (f==0.0f);
}]>;
def immFloat1 : PatLeaf<(fpimm), [{
float f = (float)N->getValueAPF().convertToFloat();
return (f==1.0f);
}]>;
def immDouble0 : PatLeaf<(fpimm), [{
double d = (double)N->getValueAPF().convertToDouble();
return (d==0.0);
}]>;
def immDouble1 : PatLeaf<(fpimm), [{
double d = (double)N->getValueAPF().convertToDouble();
return (d==1.0);
}]>;
def AS_match {
code generic = [{
return ChkMemSDNodeAddressSpace(N, llvm::ADDRESS_SPACE_GENERIC);
}];
code shared = [{
return ChkMemSDNodeAddressSpace(N, llvm::ADDRESS_SPACE_SHARED);
}];
code global = [{
return ChkMemSDNodeAddressSpace(N, llvm::ADDRESS_SPACE_GLOBAL);
}];
}
// A node that will be replaced with the current PTX version.
class PTX {
SDNodeXForm PTXVerXform = SDNodeXForm<imm, [{
return getI32Imm(Subtarget->getPTXVersion(), SDLoc(N));
}]>;
// (i32 0) will be XForm'ed to the currently used PTX version.
dag version = (PTXVerXform (i32 0));
}
def ptx : PTX;
// Generates list of n sequential register names.
// E.g. RegNames<3,"r">.ret -> ["r0", "r1", "r2" ]
class RegSeq<int n, string prefix> {
list<string> ret = !if(n, !listconcat(RegSeq<!add(n,-1), prefix>.ret,
[prefix # !add(n, -1)]),
[]);
}
//-----------------------------------
// Synchronization and shuffle functions
//-----------------------------------
let isConvergent = 1 in {
def INT_BARRIER0 : NVPTXInst<(outs), (ins),
"bar.sync \t0;",
[(int_nvvm_barrier0)]>;
def INT_BARRIERN : NVPTXInst<(outs), (ins Int32Regs:$src1),
"bar.sync \t$src1;",
[(int_nvvm_barrier_n Int32Regs:$src1)]>;
def INT_BARRIER : NVPTXInst<(outs), (ins Int32Regs:$src1, Int32Regs:$src2),
"bar.sync \t$src1, $src2;",
[(int_nvvm_barrier Int32Regs:$src1, Int32Regs:$src2)]>;
def INT_BARRIER0_POPC : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$pred),
!strconcat("{{ \n\t",
".reg .pred \t%p1; \n\t",
"setp.ne.u32 \t%p1, $pred, 0; \n\t",
"bar.red.popc.u32 \t$dst, 0, %p1; \n\t",
"}}"),
[(set Int32Regs:$dst, (int_nvvm_barrier0_popc Int32Regs:$pred))]>;
def INT_BARRIER0_AND : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$pred),
!strconcat("{{ \n\t",
".reg .pred \t%p1; \n\t",
".reg .pred \t%p2; \n\t",
"setp.ne.u32 \t%p1, $pred, 0; \n\t",
"bar.red.and.pred \t%p2, 0, %p1; \n\t",
"selp.u32 \t$dst, 1, 0, %p2; \n\t",
"}}"),
[(set Int32Regs:$dst, (int_nvvm_barrier0_and Int32Regs:$pred))]>;
def INT_BARRIER0_OR : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$pred),
!strconcat("{{ \n\t",
".reg .pred \t%p1; \n\t",
".reg .pred \t%p2; \n\t",
"setp.ne.u32 \t%p1, $pred, 0; \n\t",
"bar.red.or.pred \t%p2, 0, %p1; \n\t",
"selp.u32 \t$dst, 1, 0, %p2; \n\t",
"}}"),
[(set Int32Regs:$dst, (int_nvvm_barrier0_or Int32Regs:$pred))]>;
def INT_BAR_SYNC : NVPTXInst<(outs), (ins i32imm:$i), "bar.sync \t$i;",
[(int_nvvm_bar_sync imm:$i)]>;
def INT_BAR_WARP_SYNC_I : NVPTXInst<(outs), (ins i32imm:$i), "bar.warp.sync \t$i;",
[(int_nvvm_bar_warp_sync imm:$i)]>,
Requires<[hasPTX60, hasSM30]>;
def INT_BAR_WARP_SYNC_R : NVPTXInst<(outs), (ins Int32Regs:$i), "bar.warp.sync \t$i;",
[(int_nvvm_bar_warp_sync Int32Regs:$i)]>,
Requires<[hasPTX60, hasSM30]>;
def INT_BARRIER_SYNC_I : NVPTXInst<(outs), (ins i32imm:$i), "barrier.sync \t$i;",
[(int_nvvm_barrier_sync imm:$i)]>,
Requires<[hasPTX60, hasSM30]>;
def INT_BARRIER_SYNC_R : NVPTXInst<(outs), (ins Int32Regs:$i), "barrier.sync \t$i;",
[(int_nvvm_barrier_sync Int32Regs:$i)]>,
Requires<[hasPTX60, hasSM30]>;
def INT_BARRIER_SYNC_CNT_RR : NVPTXInst<(outs), (ins Int32Regs:$id, Int32Regs:$cnt),
"barrier.sync \t$id, $cnt;",
[(int_nvvm_barrier_sync_cnt Int32Regs:$id, Int32Regs:$cnt)]>,
Requires<[hasPTX60, hasSM30]>;
def INT_BARRIER_SYNC_CNT_RI : NVPTXInst<(outs), (ins Int32Regs:$id, i32imm:$cnt),
"barrier.sync \t$id, $cnt;",
[(int_nvvm_barrier_sync_cnt Int32Regs:$id, imm:$cnt)]>,
Requires<[hasPTX60, hasSM30]>;
def INT_BARRIER_SYNC_CNT_IR : NVPTXInst<(outs), (ins i32imm:$id, Int32Regs:$cnt),
"barrier.sync \t$id, $cnt;",
[(int_nvvm_barrier_sync_cnt imm:$id, Int32Regs:$cnt)]>,
Requires<[hasPTX60, hasSM30]>;
def INT_BARRIER_SYNC_CNT_II : NVPTXInst<(outs), (ins i32imm:$id, i32imm:$cnt),
"barrier.sync \t$id, $cnt;",
[(int_nvvm_barrier_sync_cnt imm:$id, imm:$cnt)]>,
Requires<[hasPTX60, hasSM30]>;
// shfl.{up,down,bfly,idx}.b32
multiclass SHFL<NVPTXRegClass regclass, string mode, Intrinsic IntOp> {
// The last two parameters to shfl can be regs or imms. ptxas is smart
// enough to inline constant registers, so strictly speaking we don't need to
// handle immediates here. But it's easy enough, and it makes our ptx more
// readable.
def reg : NVPTXInst<
(outs regclass:$dst),
(ins regclass:$src, Int32Regs:$offset, Int32Regs:$mask),
!strconcat("shfl.", mode, ".b32 $dst, $src, $offset, $mask;"),
[(set regclass:$dst, (IntOp regclass:$src, Int32Regs:$offset, Int32Regs:$mask))]>;
def imm1 : NVPTXInst<
(outs regclass:$dst),
(ins regclass:$src, i32imm:$offset, Int32Regs:$mask),
!strconcat("shfl.", mode, ".b32 $dst, $src, $offset, $mask;"),
[(set regclass:$dst, (IntOp regclass:$src, imm:$offset, Int32Regs:$mask))]>;
def imm2 : NVPTXInst<
(outs regclass:$dst),
(ins regclass:$src, Int32Regs:$offset, i32imm:$mask),
!strconcat("shfl.", mode, ".b32 $dst, $src, $offset, $mask;"),
[(set regclass:$dst, (IntOp regclass:$src, Int32Regs:$offset, imm:$mask))]>;
def imm3 : NVPTXInst<
(outs regclass:$dst),
(ins regclass:$src, i32imm:$offset, i32imm:$mask),
!strconcat("shfl.", mode, ".b32 $dst, $src, $offset, $mask;"),
[(set regclass:$dst, (IntOp regclass:$src, imm:$offset, imm:$mask))]>;
}
defm INT_SHFL_DOWN_I32 : SHFL<Int32Regs, "down", int_nvvm_shfl_down_i32>;
defm INT_SHFL_DOWN_F32 : SHFL<Float32Regs, "down", int_nvvm_shfl_down_f32>;
defm INT_SHFL_UP_I32 : SHFL<Int32Regs, "up", int_nvvm_shfl_up_i32>;
defm INT_SHFL_UP_F32 : SHFL<Float32Regs, "up", int_nvvm_shfl_up_f32>;
defm INT_SHFL_BFLY_I32 : SHFL<Int32Regs, "bfly", int_nvvm_shfl_bfly_i32>;
defm INT_SHFL_BFLY_F32 : SHFL<Float32Regs, "bfly", int_nvvm_shfl_bfly_f32>;
defm INT_SHFL_IDX_I32 : SHFL<Int32Regs, "idx", int_nvvm_shfl_idx_i32>;
defm INT_SHFL_IDX_F32 : SHFL<Float32Regs, "idx", int_nvvm_shfl_idx_f32>;
multiclass SHFL_SYNC<NVPTXRegClass regclass, string mode, Intrinsic IntOp> {
// Threadmask and the last two parameters to shfl.sync can be regs or imms.
// ptxas is smart enough to inline constant registers, so strictly speaking we
// don't need to handle immediates here. But it's easy enough, and it makes
// our ptx more readable.
def rrr : NVPTXInst<
(outs regclass:$dst),
(ins Int32Regs:$threadmask, regclass:$src, Int32Regs:$offset, Int32Regs:$mask),
!strconcat("shfl.sync.", mode, ".b32 $dst, $src, $offset, $mask, $threadmask;"),
[(set regclass:$dst, (IntOp Int32Regs:$threadmask, regclass:$src,
Int32Regs:$offset, Int32Regs:$mask))]>;
def rri : NVPTXInst<
(outs regclass:$dst),
(ins Int32Regs:$threadmask, regclass:$src, Int32Regs:$offset, i32imm:$mask),
!strconcat("shfl.sync.", mode, ".b32 $dst, $src, $offset, $mask, $threadmask;"),
[(set regclass:$dst, (IntOp Int32Regs:$threadmask, regclass:$src,
Int32Regs:$offset, imm:$mask))]>;
def rir : NVPTXInst<
(outs regclass:$dst),
(ins Int32Regs:$threadmask, regclass:$src, i32imm:$offset, Int32Regs:$mask),
!strconcat("shfl.sync.", mode, ".b32 $dst, $src, $offset, $mask, $threadmask;"),
[(set regclass:$dst, (IntOp Int32Regs:$threadmask, regclass:$src,
imm:$offset, Int32Regs:$mask))]>;
def rii : NVPTXInst<
(outs regclass:$dst),
(ins Int32Regs:$threadmask, regclass:$src, i32imm:$offset, i32imm:$mask),
!strconcat("shfl.sync.", mode, ".b32 $dst, $src, $offset, $mask, $threadmask;"),
[(set regclass:$dst, (IntOp Int32Regs:$threadmask, regclass:$src,
imm:$offset, imm:$mask))]>;
def irr : NVPTXInst<
(outs regclass:$dst),
(ins i32imm:$threadmask, regclass:$src, Int32Regs:$offset, Int32Regs:$mask),
!strconcat("shfl.sync.", mode, ".b32 $dst, $src, $offset, $mask, $threadmask;"),
[(set regclass:$dst, (IntOp imm:$threadmask, regclass:$src,
Int32Regs:$offset, Int32Regs:$mask))]>;
def iri : NVPTXInst<
(outs regclass:$dst),
(ins i32imm:$threadmask, regclass:$src, Int32Regs:$offset, i32imm:$mask),
!strconcat("shfl.sync.", mode, ".b32 $dst, $src, $offset, $mask, $threadmask;"),
[(set regclass:$dst, (IntOp imm:$threadmask, regclass:$src,
Int32Regs:$offset, imm:$mask))]>;
def iir : NVPTXInst<
(outs regclass:$dst),
(ins i32imm:$threadmask, regclass:$src, i32imm:$offset, Int32Regs:$mask),
!strconcat("shfl.sync.", mode, ".b32 $dst, $src, $offset, $mask, $threadmask;"),
[(set regclass:$dst, (IntOp imm:$threadmask, regclass:$src,
imm:$offset, Int32Regs:$mask))]>;
def iii : NVPTXInst<
(outs regclass:$dst),
(ins i32imm:$threadmask, regclass:$src, i32imm:$offset, i32imm:$mask),
!strconcat("shfl.sync.", mode, ".b32 $dst, $src, $offset, $mask, $threadmask;"),
[(set regclass:$dst, (IntOp imm:$threadmask, regclass:$src,
imm:$offset, imm:$mask))]>;
}
// On sm_70 these don't have to be convergent, so we may eventually want to
// implement non-convergent variant of this intrinsic.
defm INT_SHFL_SYNC_DOWN_I32 : SHFL_SYNC<Int32Regs, "down", int_nvvm_shfl_sync_down_i32>;
defm INT_SHFL_SYNC_DOWN_F32 : SHFL_SYNC<Float32Regs, "down", int_nvvm_shfl_sync_down_f32>;
defm INT_SHFL_SYNC_UP_I32 : SHFL_SYNC<Int32Regs, "up", int_nvvm_shfl_sync_up_i32>;
defm INT_SHFL_SYNC_UP_F32 : SHFL_SYNC<Float32Regs, "up", int_nvvm_shfl_sync_up_f32>;
defm INT_SHFL_SYNC_BFLY_I32 : SHFL_SYNC<Int32Regs, "bfly", int_nvvm_shfl_sync_bfly_i32>;
defm INT_SHFL_SYNC_BFLY_F32 : SHFL_SYNC<Float32Regs, "bfly", int_nvvm_shfl_sync_bfly_f32>;
defm INT_SHFL_SYNC_IDX_I32 : SHFL_SYNC<Int32Regs, "idx", int_nvvm_shfl_sync_idx_i32>;
defm INT_SHFL_SYNC_IDX_F32 : SHFL_SYNC<Float32Regs, "idx", int_nvvm_shfl_sync_idx_f32>;
// vote.{all,any,uni,ballot}
multiclass VOTE<NVPTXRegClass regclass, string mode, Intrinsic IntOp> {
def : NVPTXInst<(outs regclass:$dest), (ins Int1Regs:$pred),
"vote." # mode # " \t$dest, $pred;",
[(set regclass:$dest, (IntOp Int1Regs:$pred))]>,
Requires<[hasPTX60, hasSM30]>;
}
defm VOTE_ALL : VOTE<Int1Regs, "all.pred", int_nvvm_vote_all>;
defm VOTE_ANY : VOTE<Int1Regs, "any.pred", int_nvvm_vote_any>;
defm VOTE_UNI : VOTE<Int1Regs, "uni.pred", int_nvvm_vote_uni>;
defm VOTE_BALLOT : VOTE<Int32Regs, "ballot.b32", int_nvvm_vote_ballot>;
// vote.sync.{all,any,uni,ballot}
multiclass VOTE_SYNC<NVPTXRegClass regclass, string mode, Intrinsic IntOp> {
def i : NVPTXInst<(outs regclass:$dest), (ins i32imm:$mask, Int1Regs:$pred),
"vote.sync." # mode # " \t$dest, $pred, $mask;",
[(set regclass:$dest, (IntOp imm:$mask, Int1Regs:$pred))]>,
Requires<[hasPTX60, hasSM30]>;
def r : NVPTXInst<(outs regclass:$dest), (ins Int32Regs:$mask, Int1Regs:$pred),
"vote.sync." # mode #" \t$dest, $pred, $mask;",
[(set regclass:$dest, (IntOp Int32Regs:$mask, Int1Regs:$pred))]>,
Requires<[hasPTX60, hasSM30]>;
}
defm VOTE_SYNC_ALL : VOTE_SYNC<Int1Regs, "all.pred", int_nvvm_vote_all_sync>;
defm VOTE_SYNC_ANY : VOTE_SYNC<Int1Regs, "any.pred", int_nvvm_vote_any_sync>;
defm VOTE_SYNC_UNI : VOTE_SYNC<Int1Regs, "uni.pred", int_nvvm_vote_uni_sync>;
defm VOTE_SYNC_BALLOT : VOTE_SYNC<Int32Regs, "ballot.b32", int_nvvm_vote_ballot_sync>;
multiclass MATCH_ANY_SYNC<NVPTXRegClass regclass, string ptxtype, Intrinsic IntOp,
Operand ImmOp> {
def ii : NVPTXInst<(outs regclass:$dest), (ins i32imm:$mask, ImmOp:$value),
"match.any.sync." # ptxtype # " \t$dest, $value, $mask;",
[(set regclass:$dest, (IntOp imm:$mask, imm:$value))]>,
Requires<[hasPTX60, hasSM70]>;
def ir : NVPTXInst<(outs regclass:$dest), (ins Int32Regs:$mask, ImmOp:$value),
"match.any.sync." # ptxtype # " \t$dest, $value, $mask;",
[(set regclass:$dest, (IntOp Int32Regs:$mask, imm:$value))]>,
Requires<[hasPTX60, hasSM70]>;
def ri : NVPTXInst<(outs regclass:$dest), (ins i32imm:$mask, regclass:$value),
"match.any.sync." # ptxtype # " \t$dest, $value, $mask;",
[(set regclass:$dest, (IntOp imm:$mask, regclass:$value))]>,
Requires<[hasPTX60, hasSM70]>;
def rr : NVPTXInst<(outs regclass:$dest), (ins Int32Regs:$mask, regclass:$value),
"match.any.sync." # ptxtype # " \t$dest, $value, $mask;",
[(set regclass:$dest, (IntOp Int32Regs:$mask, regclass:$value))]>,
Requires<[hasPTX60, hasSM70]>;
}
defm MATCH_ANY_SYNC_32 : MATCH_ANY_SYNC<Int32Regs, "b32", int_nvvm_match_any_sync_i32,
i32imm>;
defm MATCH_ANY_SYNC_64 : MATCH_ANY_SYNC<Int64Regs, "b64", int_nvvm_match_any_sync_i64,
i64imm>;
multiclass MATCH_ALLP_SYNC<NVPTXRegClass regclass, string ptxtype, Intrinsic IntOp,
Operand ImmOp> {
def ii : NVPTXInst<(outs regclass:$dest, Int1Regs:$pred),
(ins i32imm:$mask, ImmOp:$value),
"match.all.sync." # ptxtype # " \t$dest|$pred, $value, $mask;",
[(set regclass:$dest, Int1Regs:$pred, (IntOp imm:$mask, imm:$value))]>,
Requires<[hasPTX60, hasSM70]>;
def ir : NVPTXInst<(outs regclass:$dest, Int1Regs:$pred),
(ins Int32Regs:$mask, ImmOp:$value),
"match.all.sync." # ptxtype # " \t$dest|$pred, $value, $mask;",
[(set regclass:$dest, Int1Regs:$pred, (IntOp Int32Regs:$mask, imm:$value))]>,
Requires<[hasPTX60, hasSM70]>;
def ri : NVPTXInst<(outs regclass:$dest, Int1Regs:$pred),
(ins i32imm:$mask, regclass:$value),
"match.all.sync." # ptxtype # " \t$dest|$pred, $value, $mask;",
[(set regclass:$dest, Int1Regs:$pred, (IntOp imm:$mask, regclass:$value))]>,
Requires<[hasPTX60, hasSM70]>;
def rr : NVPTXInst<(outs regclass:$dest, Int1Regs:$pred),
(ins Int32Regs:$mask, regclass:$value),
"match.all.sync." # ptxtype # " \t$dest|$pred, $value, $mask;",
[(set regclass:$dest, Int1Regs:$pred, (IntOp Int32Regs:$mask, regclass:$value))]>,
Requires<[hasPTX60, hasSM70]>;
}
defm MATCH_ALLP_SYNC_32 : MATCH_ALLP_SYNC<Int32Regs, "b32", int_nvvm_match_all_sync_i32p,
i32imm>;
defm MATCH_ALLP_SYNC_64 : MATCH_ALLP_SYNC<Int64Regs, "b64", int_nvvm_match_all_sync_i64p,
i64imm>;
} // isConvergent = 1
//-----------------------------------
// Explicit Memory Fence Functions
//-----------------------------------
class MEMBAR<string StrOp, Intrinsic IntOP> :
NVPTXInst<(outs), (ins),
StrOp, [(IntOP)]>;
def INT_MEMBAR_CTA : MEMBAR<"membar.cta;", int_nvvm_membar_cta>;
def INT_MEMBAR_GL : MEMBAR<"membar.gl;", int_nvvm_membar_gl>;
def INT_MEMBAR_SYS : MEMBAR<"membar.sys;", int_nvvm_membar_sys>;
//-----------------------------------
// Math Functions
//-----------------------------------
// Map min(1.0, max(0.0, x)) to sat(x)
// Note that max(0.0, min(x, 1.0)) cannot be mapped to sat(x) because when x is
// NaN
// max(0.0, min(x, 1.0)) is 1.0 while sat(x) is 0.
// Same story for fmax, fmin.
def : Pat<(int_nvvm_fmin_f immFloat1,
(int_nvvm_fmax_f immFloat0, Float32Regs:$a)),
(CVT_f32_f32 Float32Regs:$a, CvtSAT)>;
def : Pat<(int_nvvm_fmin_f immFloat1,
(int_nvvm_fmax_f Float32Regs:$a, immFloat0)),
(CVT_f32_f32 Float32Regs:$a, CvtSAT)>;
def : Pat<(int_nvvm_fmin_f
(int_nvvm_fmax_f immFloat0, Float32Regs:$a), immFloat1),
(CVT_f32_f32 Float32Regs:$a, CvtSAT)>;
def : Pat<(int_nvvm_fmin_f
(int_nvvm_fmax_f Float32Regs:$a, immFloat0), immFloat1),
(CVT_f32_f32 Float32Regs:$a, CvtSAT)>;
def : Pat<(int_nvvm_fmin_d immDouble1,
(int_nvvm_fmax_d immDouble0, Float64Regs:$a)),
(CVT_f64_f64 Float64Regs:$a, CvtSAT)>;
def : Pat<(int_nvvm_fmin_d immDouble1,
(int_nvvm_fmax_d Float64Regs:$a, immDouble0)),
(CVT_f64_f64 Float64Regs:$a, CvtSAT)>;
def : Pat<(int_nvvm_fmin_d
(int_nvvm_fmax_d immDouble0, Float64Regs:$a), immDouble1),
(CVT_f64_f64 Float64Regs:$a, CvtSAT)>;
def : Pat<(int_nvvm_fmin_d
(int_nvvm_fmax_d Float64Regs:$a, immDouble0), immDouble1),
(CVT_f64_f64 Float64Regs:$a, CvtSAT)>;
// We need a full string for OpcStr here because we need to deal with case like
// INT_PTX_RECIP.
class F_MATH_1<string OpcStr, NVPTXRegClass target_regclass,
NVPTXRegClass src_regclass, Intrinsic IntOP>
: NVPTXInst<(outs target_regclass:$dst), (ins src_regclass:$src0),
OpcStr,
[(set target_regclass:$dst, (IntOP src_regclass:$src0))]>;
// We need a full string for OpcStr here because we need to deal with the case
// like INT_PTX_NATIVE_POWR_F.
class F_MATH_2<string OpcStr, NVPTXRegClass t_regclass,
NVPTXRegClass s0_regclass, NVPTXRegClass s1_regclass, Intrinsic IntOP>
: NVPTXInst<(outs t_regclass:$dst),
(ins s0_regclass:$src0, s1_regclass:$src1),
OpcStr,
[(set t_regclass:$dst, (IntOP s0_regclass:$src0, s1_regclass:$src1))]>;
class F_MATH_3<string OpcStr, NVPTXRegClass t_regclass,
NVPTXRegClass s0_regclass, NVPTXRegClass s1_regclass,
NVPTXRegClass s2_regclass, Intrinsic IntOP>
: NVPTXInst<(outs t_regclass:$dst),
(ins s0_regclass:$src0, s1_regclass:$src1, s2_regclass:$src2),
OpcStr,
[(set t_regclass:$dst,
(IntOP s0_regclass:$src0, s1_regclass:$src1, s2_regclass:$src2))]>;
//
// MISC
//
def INT_NVVM_PRMT : F_MATH_3<"prmt.b32 \t$dst, $src0, $src1, $src2;", Int32Regs,
Int32Regs, Int32Regs, Int32Regs, int_nvvm_prmt>;
//
// Min Max
//
def INT_NVVM_FMIN_F : F_MATH_2<"min.f32 \t$dst, $src0, $src1;", Float32Regs,
Float32Regs, Float32Regs, int_nvvm_fmin_f>;
def INT_NVVM_FMIN_FTZ_F : F_MATH_2<"min.ftz.f32 \t$dst, $src0, $src1;",
Float32Regs, Float32Regs, Float32Regs, int_nvvm_fmin_ftz_f>;
def INT_NVVM_FMAX_F : F_MATH_2<"max.f32 \t$dst, $src0, $src1;", Float32Regs,
Float32Regs, Float32Regs, int_nvvm_fmax_f>;
def INT_NVVM_FMAX_FTZ_F : F_MATH_2<"max.ftz.f32 \t$dst, $src0, $src1;",
Float32Regs, Float32Regs, Float32Regs, int_nvvm_fmax_ftz_f>;
def INT_NVVM_FMIN_D : F_MATH_2<"min.f64 \t$dst, $src0, $src1;", Float64Regs,
Float64Regs, Float64Regs, int_nvvm_fmin_d>;
def INT_NVVM_FMAX_D : F_MATH_2<"max.f64 \t$dst, $src0, $src1;", Float64Regs,
Float64Regs, Float64Regs, int_nvvm_fmax_d>;
//
// Multiplication
//
def INT_NVVM_MULHI_I : F_MATH_2<"mul.hi.s32 \t$dst, $src0, $src1;", Int32Regs,
Int32Regs, Int32Regs, int_nvvm_mulhi_i>;
def INT_NVVM_MULHI_UI : F_MATH_2<"mul.hi.u32 \t$dst, $src0, $src1;", Int32Regs,
Int32Regs, Int32Regs, int_nvvm_mulhi_ui>;
def INT_NVVM_MULHI_LL : F_MATH_2<"mul.hi.s64 \t$dst, $src0, $src1;", Int64Regs,
Int64Regs, Int64Regs, int_nvvm_mulhi_ll>;
def INT_NVVM_MULHI_ULL : F_MATH_2<"mul.hi.u64 \t$dst, $src0, $src1;", Int64Regs,
Int64Regs, Int64Regs, int_nvvm_mulhi_ull>;
def INT_NVVM_MUL_RN_FTZ_F : F_MATH_2<"mul.rn.ftz.f32 \t$dst, $src0, $src1;",
Float32Regs, Float32Regs, Float32Regs, int_nvvm_mul_rn_ftz_f>;
def INT_NVVM_MUL_RN_F : F_MATH_2<"mul.rn.f32 \t$dst, $src0, $src1;",
Float32Regs, Float32Regs, Float32Regs, int_nvvm_mul_rn_f>;
def INT_NVVM_MUL_RZ_FTZ_F : F_MATH_2<"mul.rz.ftz.f32 \t$dst, $src0, $src1;",
Float32Regs, Float32Regs, Float32Regs, int_nvvm_mul_rz_ftz_f>;
def INT_NVVM_MUL_RZ_F : F_MATH_2<"mul.rz.f32 \t$dst, $src0, $src1;",
Float32Regs, Float32Regs, Float32Regs, int_nvvm_mul_rz_f>;
def INT_NVVM_MUL_RM_FTZ_F : F_MATH_2<"mul.rm.ftz.f32 \t$dst, $src0, $src1;",
Float32Regs, Float32Regs, Float32Regs, int_nvvm_mul_rm_ftz_f>;
def INT_NVVM_MUL_RM_F : F_MATH_2<"mul.rm.f32 \t$dst, $src0, $src1;",
Float32Regs, Float32Regs, Float32Regs, int_nvvm_mul_rm_f>;
def INT_NVVM_MUL_RP_FTZ_F : F_MATH_2<"mul.rp.ftz.f32 \t$dst, $src0, $src1;",
Float32Regs, Float32Regs, Float32Regs, int_nvvm_mul_rp_ftz_f>;
def INT_NVVM_MUL_RP_F : F_MATH_2<"mul.rp.f32 \t$dst, $src0, $src1;",
Float32Regs, Float32Regs, Float32Regs, int_nvvm_mul_rp_f>;
def INT_NVVM_MUL_RN_D : F_MATH_2<"mul.rn.f64 \t$dst, $src0, $src1;",
Float64Regs, Float64Regs, Float64Regs, int_nvvm_mul_rn_d>;
def INT_NVVM_MUL_RZ_D : F_MATH_2<"mul.rz.f64 \t$dst, $src0, $src1;",
Float64Regs, Float64Regs, Float64Regs, int_nvvm_mul_rz_d>;
def INT_NVVM_MUL_RM_D : F_MATH_2<"mul.rm.f64 \t$dst, $src0, $src1;",
Float64Regs, Float64Regs, Float64Regs, int_nvvm_mul_rm_d>;
def INT_NVVM_MUL_RP_D : F_MATH_2<"mul.rp.f64 \t$dst, $src0, $src1;",
Float64Regs, Float64Regs, Float64Regs, int_nvvm_mul_rp_d>;
def INT_NVVM_MUL24_I : F_MATH_2<"mul24.lo.s32 \t$dst, $src0, $src1;",
Int32Regs, Int32Regs, Int32Regs, int_nvvm_mul24_i>;
def INT_NVVM_MUL24_UI : F_MATH_2<"mul24.lo.u32 \t$dst, $src0, $src1;",
Int32Regs, Int32Regs, Int32Regs, int_nvvm_mul24_ui>;
//
// Div
//
def INT_NVVM_DIV_APPROX_FTZ_F
: F_MATH_2<"div.approx.ftz.f32 \t$dst, $src0, $src1;", Float32Regs,
Float32Regs, Float32Regs, int_nvvm_div_approx_ftz_f>;
def INT_NVVM_DIV_APPROX_F : F_MATH_2<"div.approx.f32 \t$dst, $src0, $src1;",
Float32Regs, Float32Regs, Float32Regs, int_nvvm_div_approx_f>;
def INT_NVVM_DIV_RN_FTZ_F : F_MATH_2<"div.rn.ftz.f32 \t$dst, $src0, $src1;",
Float32Regs, Float32Regs, Float32Regs, int_nvvm_div_rn_ftz_f>;
def INT_NVVM_DIV_RN_F : F_MATH_2<"div.rn.f32 \t$dst, $src0, $src1;",
Float32Regs, Float32Regs, Float32Regs, int_nvvm_div_rn_f>;
def INT_NVVM_DIV_RZ_FTZ_F : F_MATH_2<"div.rz.ftz.f32 \t$dst, $src0, $src1;",
Float32Regs, Float32Regs, Float32Regs, int_nvvm_div_rz_ftz_f>;
def INT_NVVM_DIV_RZ_F : F_MATH_2<"div.rz.f32 \t$dst, $src0, $src1;",
Float32Regs, Float32Regs, Float32Regs, int_nvvm_div_rz_f>;
def INT_NVVM_DIV_RM_FTZ_F : F_MATH_2<"div.rm.ftz.f32 \t$dst, $src0, $src1;",
Float32Regs, Float32Regs, Float32Regs, int_nvvm_div_rm_ftz_f>;
def INT_NVVM_DIV_RM_F : F_MATH_2<"div.rm.f32 \t$dst, $src0, $src1;",
Float32Regs, Float32Regs, Float32Regs, int_nvvm_div_rm_f>;
def INT_NVVM_DIV_RP_FTZ_F : F_MATH_2<"div.rp.ftz.f32 \t$dst, $src0, $src1;",
Float32Regs, Float32Regs, Float32Regs, int_nvvm_div_rp_ftz_f>;
def INT_NVVM_DIV_RP_F : F_MATH_2<"div.rp.f32 \t$dst, $src0, $src1;",
Float32Regs, Float32Regs, Float32Regs, int_nvvm_div_rp_f>;
def INT_NVVM_DIV_RN_D : F_MATH_2<"div.rn.f64 \t$dst, $src0, $src1;",
Float64Regs, Float64Regs, Float64Regs, int_nvvm_div_rn_d>;
def INT_NVVM_DIV_RZ_D : F_MATH_2<"div.rz.f64 \t$dst, $src0, $src1;",
Float64Regs, Float64Regs, Float64Regs, int_nvvm_div_rz_d>;
def INT_NVVM_DIV_RM_D : F_MATH_2<"div.rm.f64 \t$dst, $src0, $src1;",
Float64Regs, Float64Regs, Float64Regs, int_nvvm_div_rm_d>;
def INT_NVVM_DIV_RP_D : F_MATH_2<"div.rp.f64 \t$dst, $src0, $src1;",
Float64Regs, Float64Regs, Float64Regs, int_nvvm_div_rp_d>;
//
// Sad
//
def INT_NVVM_SAD_I : F_MATH_3<"sad.s32 \t$dst, $src0, $src1, $src2;",
Int32Regs, Int32Regs, Int32Regs, Int32Regs, int_nvvm_sad_i>;
def INT_NVVM_SAD_UI : F_MATH_3<"sad.u32 \t$dst, $src0, $src1, $src2;",
Int32Regs, Int32Regs, Int32Regs, Int32Regs, int_nvvm_sad_ui>;
//
// Floor Ceil
//
def : Pat<(int_nvvm_floor_ftz_f Float32Regs:$a),
(CVT_f32_f32 Float32Regs:$a, CvtRMI_FTZ)>;
def : Pat<(int_nvvm_floor_f Float32Regs:$a),
(CVT_f32_f32 Float32Regs:$a, CvtRMI)>;
def : Pat<(int_nvvm_floor_d Float64Regs:$a),
(CVT_f64_f64 Float64Regs:$a, CvtRMI)>;
def : Pat<(int_nvvm_ceil_ftz_f Float32Regs:$a),
(CVT_f32_f32 Float32Regs:$a, CvtRPI_FTZ)>;
def : Pat<(int_nvvm_ceil_f Float32Regs:$a),
(CVT_f32_f32 Float32Regs:$a, CvtRPI)>;
def : Pat<(int_nvvm_ceil_d Float64Regs:$a),
(CVT_f64_f64 Float64Regs:$a, CvtRPI)>;
//
// Abs
//
def INT_NVVM_FABS_FTZ_F : F_MATH_1<"abs.ftz.f32 \t$dst, $src0;", Float32Regs,
Float32Regs, int_nvvm_fabs_ftz_f>;
def INT_NVVM_FABS_F : F_MATH_1<"abs.f32 \t$dst, $src0;", Float32Regs,
Float32Regs, int_nvvm_fabs_f>;
def INT_NVVM_FABS_D : F_MATH_1<"abs.f64 \t$dst, $src0;", Float64Regs,
Float64Regs, int_nvvm_fabs_d>;
//
// Round
//
def : Pat<(int_nvvm_round_ftz_f Float32Regs:$a),
(CVT_f32_f32 Float32Regs:$a, CvtRNI_FTZ)>;
def : Pat<(int_nvvm_round_f Float32Regs:$a),
(CVT_f32_f32 Float32Regs:$a, CvtRNI)>;
def : Pat<(int_nvvm_round_d Float64Regs:$a),
(CVT_f64_f64 Float64Regs:$a, CvtRNI)>;
//
// Trunc
//
def : Pat<(int_nvvm_trunc_ftz_f Float32Regs:$a),
(CVT_f32_f32 Float32Regs:$a, CvtRZI_FTZ)>;
def : Pat<(int_nvvm_trunc_f Float32Regs:$a),
(CVT_f32_f32 Float32Regs:$a, CvtRZI)>;
def : Pat<(int_nvvm_trunc_d Float64Regs:$a),
(CVT_f64_f64 Float64Regs:$a, CvtRZI)>;
//
// Saturate
//
def : Pat<(int_nvvm_saturate_ftz_f Float32Regs:$a),
(CVT_f32_f32 Float32Regs:$a, CvtSAT_FTZ)>;
def : Pat<(int_nvvm_saturate_f Float32Regs:$a),
(CVT_f32_f32 Float32Regs:$a, CvtSAT)>;
def : Pat<(int_nvvm_saturate_d Float64Regs:$a),
(CVT_f64_f64 Float64Regs:$a, CvtSAT)>;
//
// Exp2 Log2
//
def INT_NVVM_EX2_APPROX_FTZ_F : F_MATH_1<"ex2.approx.ftz.f32 \t$dst, $src0;",
Float32Regs, Float32Regs, int_nvvm_ex2_approx_ftz_f>;
def INT_NVVM_EX2_APPROX_F : F_MATH_1<"ex2.approx.f32 \t$dst, $src0;",
Float32Regs, Float32Regs, int_nvvm_ex2_approx_f>;
def INT_NVVM_EX2_APPROX_D : F_MATH_1<"ex2.approx.f64 \t$dst, $src0;",
Float64Regs, Float64Regs, int_nvvm_ex2_approx_d>;
def INT_NVVM_LG2_APPROX_FTZ_F : F_MATH_1<"lg2.approx.ftz.f32 \t$dst, $src0;",
Float32Regs, Float32Regs, int_nvvm_lg2_approx_ftz_f>;
def INT_NVVM_LG2_APPROX_F : F_MATH_1<"lg2.approx.f32 \t$dst, $src0;",
Float32Regs, Float32Regs, int_nvvm_lg2_approx_f>;
def INT_NVVM_LG2_APPROX_D : F_MATH_1<"lg2.approx.f64 \t$dst, $src0;",
Float64Regs, Float64Regs, int_nvvm_lg2_approx_d>;
//
// Sin Cos
//
def INT_NVVM_SIN_APPROX_FTZ_F : F_MATH_1<"sin.approx.ftz.f32 \t$dst, $src0;",
Float32Regs, Float32Regs, int_nvvm_sin_approx_ftz_f>;
def INT_NVVM_SIN_APPROX_F : F_MATH_1<"sin.approx.f32 \t$dst, $src0;",
Float32Regs, Float32Regs, int_nvvm_sin_approx_f>;
def INT_NVVM_COS_APPROX_FTZ_F : F_MATH_1<"cos.approx.ftz.f32 \t$dst, $src0;",
Float32Regs, Float32Regs, int_nvvm_cos_approx_ftz_f>;
def INT_NVVM_COS_APPROX_F : F_MATH_1<"cos.approx.f32 \t$dst, $src0;",
Float32Regs, Float32Regs, int_nvvm_cos_approx_f>;
//
// Fma
//
def INT_NVVM_FMA_RN_FTZ_F
: F_MATH_3<"fma.rn.ftz.f32 \t$dst, $src0, $src1, $src2;", Float32Regs,
Float32Regs, Float32Regs, Float32Regs, int_nvvm_fma_rn_ftz_f>;
def INT_NVVM_FMA_RN_F : F_MATH_3<"fma.rn.f32 \t$dst, $src0, $src1, $src2;",
Float32Regs, Float32Regs, Float32Regs, Float32Regs, int_nvvm_fma_rn_f>;
def INT_NVVM_FMA_RZ_FTZ_F
: F_MATH_3<"fma.rz.ftz.f32 \t$dst, $src0, $src1, $src2;", Float32Regs,
Float32Regs, Float32Regs, Float32Regs, int_nvvm_fma_rz_ftz_f>;
def INT_NVVM_FMA_RZ_F : F_MATH_3<"fma.rz.f32 \t$dst, $src0, $src1, $src2;",
Float32Regs, Float32Regs, Float32Regs, Float32Regs, int_nvvm_fma_rz_f>;
def INT_NVVM_FMA_RM_FTZ_F
: F_MATH_3<"fma.rm.ftz.f32 \t$dst, $src0, $src1, $src2;", Float32Regs,
Float32Regs, Float32Regs, Float32Regs, int_nvvm_fma_rm_ftz_f>;
def INT_NVVM_FMA_RM_F : F_MATH_3<"fma.rm.f32 \t$dst, $src0, $src1, $src2;",
Float32Regs, Float32Regs, Float32Regs, Float32Regs, int_nvvm_fma_rm_f>;
def INT_NVVM_FMA_RP_FTZ_F
: F_MATH_3<"fma.rp.ftz.f32 \t$dst, $src0, $src1, $src2;", Float32Regs,
Float32Regs, Float32Regs, Float32Regs, int_nvvm_fma_rp_ftz_f>;
def INT_NVVM_FMA_RP_F : F_MATH_3<"fma.rp.f32 \t$dst, $src0, $src1, $src2;",
Float32Regs, Float32Regs, Float32Regs, Float32Regs, int_nvvm_fma_rp_f>;
def INT_NVVM_FMA_RN_D : F_MATH_3<"fma.rn.f64 \t$dst, $src0, $src1, $src2;",
Float64Regs, Float64Regs, Float64Regs, Float64Regs, int_nvvm_fma_rn_d>;
def INT_NVVM_FMA_RZ_D : F_MATH_3<"fma.rz.f64 \t$dst, $src0, $src1, $src2;",
Float64Regs, Float64Regs, Float64Regs, Float64Regs, int_nvvm_fma_rz_d>;
def INT_NVVM_FMA_RM_D : F_MATH_3<"fma.rm.f64 \t$dst, $src0, $src1, $src2;",
Float64Regs, Float64Regs, Float64Regs, Float64Regs, int_nvvm_fma_rm_d>;
def INT_NVVM_FMA_RP_D : F_MATH_3<"fma.rp.f64 \t$dst, $src0, $src1, $src2;",
Float64Regs, Float64Regs, Float64Regs, Float64Regs, int_nvvm_fma_rp_d>;
//
// Rcp
//
def INT_NVVM_RCP_RN_FTZ_F : F_MATH_1<"rcp.rn.ftz.f32 \t$dst, $src0;",
Float32Regs, Float32Regs, int_nvvm_rcp_rn_ftz_f>;
def INT_NVVM_RCP_RN_F : F_MATH_1<"rcp.rn.f32 \t$dst, $src0;",
Float32Regs, Float32Regs, int_nvvm_rcp_rn_f>;
def INT_NVVM_RCP_RZ_FTZ_F : F_MATH_1<"rcp.rz.ftz.f32 \t$dst, $src0;",
Float32Regs, Float32Regs, int_nvvm_rcp_rz_ftz_f>;
def INT_NVVM_RCP_RZ_F : F_MATH_1<"rcp.rz.f32 \t$dst, $src0;",
Float32Regs, Float32Regs, int_nvvm_rcp_rz_f>;
def INT_NVVM_RCP_RM_FTZ_F : F_MATH_1<"rcp.rm.ftz.f32 \t$dst, $src0;",
Float32Regs, Float32Regs, int_nvvm_rcp_rm_ftz_f>;
def INT_NVVM_RCP_RM_F : F_MATH_1<"rcp.rm.f32 \t$dst, $src0;",
Float32Regs, Float32Regs, int_nvvm_rcp_rm_f>;
def INT_NVVM_RCP_RP_FTZ_F : F_MATH_1<"rcp.rp.ftz.f32 \t$dst, $src0;",
Float32Regs, Float32Regs, int_nvvm_rcp_rp_ftz_f>;
def INT_NVVM_RCP_RP_F : F_MATH_1<"rcp.rp.f32 \t$dst, $src0;",
Float32Regs, Float32Regs, int_nvvm_rcp_rp_f>;
def INT_NVVM_RCP_RN_D : F_MATH_1<"rcp.rn.f64 \t$dst, $src0;", Float64Regs,
Float64Regs, int_nvvm_rcp_rn_d>;
def INT_NVVM_RCP_RZ_D : F_MATH_1<"rcp.rz.f64 \t$dst, $src0;", Float64Regs,
Float64Regs, int_nvvm_rcp_rz_d>;
def INT_NVVM_RCP_RM_D : F_MATH_1<"rcp.rm.f64 \t$dst, $src0;", Float64Regs,
Float64Regs, int_nvvm_rcp_rm_d>;
def INT_NVVM_RCP_RP_D : F_MATH_1<"rcp.rp.f64 \t$dst, $src0;", Float64Regs,
Float64Regs, int_nvvm_rcp_rp_d>;
def INT_NVVM_RCP_APPROX_FTZ_D : F_MATH_1<"rcp.approx.ftz.f64 \t$dst, $src0;",
Float64Regs, Float64Regs, int_nvvm_rcp_approx_ftz_d>;
//
// Sqrt
//
def INT_NVVM_SQRT_RN_FTZ_F : F_MATH_1<"sqrt.rn.ftz.f32 \t$dst, $src0;",
Float32Regs, Float32Regs, int_nvvm_sqrt_rn_ftz_f>;
def INT_NVVM_SQRT_RN_F : F_MATH_1<"sqrt.rn.f32 \t$dst, $src0;", Float32Regs,
Float32Regs, int_nvvm_sqrt_rn_f>;
def INT_NVVM_SQRT_RZ_FTZ_F : F_MATH_1<"sqrt.rz.ftz.f32 \t$dst, $src0;",
Float32Regs, Float32Regs, int_nvvm_sqrt_rz_ftz_f>;
def INT_NVVM_SQRT_RZ_F : F_MATH_1<"sqrt.rz.f32 \t$dst, $src0;", Float32Regs,
Float32Regs, int_nvvm_sqrt_rz_f>;
def INT_NVVM_SQRT_RM_FTZ_F : F_MATH_1<"sqrt.rm.ftz.f32 \t$dst, $src0;",
Float32Regs, Float32Regs, int_nvvm_sqrt_rm_ftz_f>;
def INT_NVVM_SQRT_RM_F : F_MATH_1<"sqrt.rm.f32 \t$dst, $src0;", Float32Regs,
Float32Regs, int_nvvm_sqrt_rm_f>;
def INT_NVVM_SQRT_RP_FTZ_F : F_MATH_1<"sqrt.rp.ftz.f32 \t$dst, $src0;",
Float32Regs, Float32Regs, int_nvvm_sqrt_rp_ftz_f>;
def INT_NVVM_SQRT_RP_F : F_MATH_1<"sqrt.rp.f32 \t$dst, $src0;", Float32Regs,
Float32Regs, int_nvvm_sqrt_rp_f>;
def INT_NVVM_SQRT_APPROX_FTZ_F : F_MATH_1<"sqrt.approx.ftz.f32 \t$dst, $src0;",
Float32Regs, Float32Regs, int_nvvm_sqrt_approx_ftz_f>;
def INT_NVVM_SQRT_APPROX_F : F_MATH_1<"sqrt.approx.f32 \t$dst, $src0;",
Float32Regs, Float32Regs, int_nvvm_sqrt_approx_f>;
def INT_NVVM_SQRT_RN_D : F_MATH_1<"sqrt.rn.f64 \t$dst, $src0;", Float64Regs,
Float64Regs, int_nvvm_sqrt_rn_d>;
def INT_NVVM_SQRT_RZ_D : F_MATH_1<"sqrt.rz.f64 \t$dst, $src0;", Float64Regs,
Float64Regs, int_nvvm_sqrt_rz_d>;
def INT_NVVM_SQRT_RM_D : F_MATH_1<"sqrt.rm.f64 \t$dst, $src0;", Float64Regs,
Float64Regs, int_nvvm_sqrt_rm_d>;
def INT_NVVM_SQRT_RP_D : F_MATH_1<"sqrt.rp.f64 \t$dst, $src0;", Float64Regs,
Float64Regs, int_nvvm_sqrt_rp_d>;
// nvvm_sqrt intrinsic
def : Pat<(int_nvvm_sqrt_f Float32Regs:$a),
(INT_NVVM_SQRT_RN_FTZ_F Float32Regs:$a)>, Requires<[doF32FTZ, do_SQRTF32_RN]>;
def : Pat<(int_nvvm_sqrt_f Float32Regs:$a),
(INT_NVVM_SQRT_RN_F Float32Regs:$a)>, Requires<[do_SQRTF32_RN]>;
def : Pat<(int_nvvm_sqrt_f Float32Regs:$a),
(INT_NVVM_SQRT_APPROX_FTZ_F Float32Regs:$a)>, Requires<[doF32FTZ]>;
def : Pat<(int_nvvm_sqrt_f Float32Regs:$a),
(INT_NVVM_SQRT_APPROX_F Float32Regs:$a)>;
//
// Rsqrt
//
def INT_NVVM_RSQRT_APPROX_FTZ_F
: F_MATH_1<"rsqrt.approx.ftz.f32 \t$dst, $src0;", Float32Regs, Float32Regs,
int_nvvm_rsqrt_approx_ftz_f>;
def INT_NVVM_RSQRT_APPROX_F : F_MATH_1<"rsqrt.approx.f32 \t$dst, $src0;",
Float32Regs, Float32Regs, int_nvvm_rsqrt_approx_f>;
def INT_NVVM_RSQRT_APPROX_D : F_MATH_1<"rsqrt.approx.f64 \t$dst, $src0;",
Float64Regs, Float64Regs, int_nvvm_rsqrt_approx_d>;
//
// Add
//
def INT_NVVM_ADD_RN_FTZ_F : F_MATH_2<"add.rn.ftz.f32 \t$dst, $src0, $src1;",
Float32Regs, Float32Regs, Float32Regs, int_nvvm_add_rn_ftz_f>;
def INT_NVVM_ADD_RN_F : F_MATH_2<"add.rn.f32 \t$dst, $src0, $src1;",
Float32Regs, Float32Regs, Float32Regs, int_nvvm_add_rn_f>;
def INT_NVVM_ADD_RZ_FTZ_F : F_MATH_2<"add.rz.ftz.f32 \t$dst, $src0, $src1;",
Float32Regs, Float32Regs, Float32Regs, int_nvvm_add_rz_ftz_f>;
def INT_NVVM_ADD_RZ_F : F_MATH_2<"add.rz.f32 \t$dst, $src0, $src1;",
Float32Regs, Float32Regs, Float32Regs, int_nvvm_add_rz_f>;
def INT_NVVM_ADD_RM_FTZ_F : F_MATH_2<"add.rm.ftz.f32 \t$dst, $src0, $src1;",
Float32Regs, Float32Regs, Float32Regs, int_nvvm_add_rm_ftz_f>;
def INT_NVVM_ADD_RM_F : F_MATH_2<"add.rm.f32 \t$dst, $src0, $src1;",
Float32Regs, Float32Regs, Float32Regs, int_nvvm_add_rm_f>;
def INT_NVVM_ADD_RP_FTZ_F : F_MATH_2<"add.rp.ftz.f32 \t$dst, $src0, $src1;",
Float32Regs, Float32Regs, Float32Regs, int_nvvm_add_rp_ftz_f>;
def INT_NVVM_ADD_RP_F : F_MATH_2<"add.rp.f32 \t$dst, $src0, $src1;",
Float32Regs, Float32Regs, Float32Regs, int_nvvm_add_rp_f>;
def INT_NVVM_ADD_RN_D : F_MATH_2<"add.rn.f64 \t$dst, $src0, $src1;",
Float64Regs, Float64Regs, Float64Regs, int_nvvm_add_rn_d>;
def INT_NVVM_ADD_RZ_D : F_MATH_2<"add.rz.f64 \t$dst, $src0, $src1;",
Float64Regs, Float64Regs, Float64Regs, int_nvvm_add_rz_d>;
def INT_NVVM_ADD_RM_D : F_MATH_2<"add.rm.f64 \t$dst, $src0, $src1;",
Float64Regs, Float64Regs, Float64Regs, int_nvvm_add_rm_d>;
def INT_NVVM_ADD_RP_D : F_MATH_2<"add.rp.f64 \t$dst, $src0, $src1;",
Float64Regs, Float64Regs, Float64Regs, int_nvvm_add_rp_d>;
//
// Convert
//
def : Pat<(int_nvvm_d2f_rn_ftz Float64Regs:$a),
(CVT_f32_f64 Float64Regs:$a, CvtRN_FTZ)>;
def : Pat<(int_nvvm_d2f_rn Float64Regs:$a),
(CVT_f32_f64 Float64Regs:$a, CvtRN)>;
def : Pat<(int_nvvm_d2f_rz_ftz Float64Regs:$a),
(CVT_f32_f64 Float64Regs:$a, CvtRZ_FTZ)>;
def : Pat<(int_nvvm_d2f_rz Float64Regs:$a),
(CVT_f32_f64 Float64Regs:$a, CvtRZ)>;
def : Pat<(int_nvvm_d2f_rm_ftz Float64Regs:$a),
(CVT_f32_f64 Float64Regs:$a, CvtRM_FTZ)>;
def : Pat<(int_nvvm_d2f_rm Float64Regs:$a),
(CVT_f32_f64 Float64Regs:$a, CvtRM)>;
def : Pat<(int_nvvm_d2f_rp_ftz Float64Regs:$a),
(CVT_f32_f64 Float64Regs:$a, CvtRP_FTZ)>;
def : Pat<(int_nvvm_d2f_rp Float64Regs:$a),
(CVT_f32_f64 Float64Regs:$a, CvtRP)>;
def : Pat<(int_nvvm_d2i_rn Float64Regs:$a),
(CVT_s32_f64 Float64Regs:$a, CvtRNI)>;
def : Pat<(int_nvvm_d2i_rz Float64Regs:$a),
(CVT_s32_f64 Float64Regs:$a, CvtRZI)>;
def : Pat<(int_nvvm_d2i_rm Float64Regs:$a),
(CVT_s32_f64 Float64Regs:$a, CvtRMI)>;
def : Pat<(int_nvvm_d2i_rp Float64Regs:$a),
(CVT_s32_f64 Float64Regs:$a, CvtRPI)>;
def : Pat<(int_nvvm_d2ui_rn Float64Regs:$a),
(CVT_u32_f64 Float64Regs:$a, CvtRNI)>;
def : Pat<(int_nvvm_d2ui_rz Float64Regs:$a),
(CVT_u32_f64 Float64Regs:$a, CvtRZI)>;
def : Pat<(int_nvvm_d2ui_rm Float64Regs:$a),
(CVT_u32_f64 Float64Regs:$a, CvtRMI)>;
def : Pat<(int_nvvm_d2ui_rp Float64Regs:$a),
(CVT_u32_f64 Float64Regs:$a, CvtRPI)>;
def : Pat<(int_nvvm_i2d_rn Int32Regs:$a),
(CVT_f64_s32 Int32Regs:$a, CvtRN)>;
def : Pat<(int_nvvm_i2d_rz Int32Regs:$a),
(CVT_f64_s32 Int32Regs:$a, CvtRZ)>;
def : Pat<(int_nvvm_i2d_rm Int32Regs:$a),
(CVT_f64_s32 Int32Regs:$a, CvtRM)>;
def : Pat<(int_nvvm_i2d_rp Int32Regs:$a),
(CVT_f64_s32 Int32Regs:$a, CvtRP)>;
def : Pat<(int_nvvm_ui2d_rn Int32Regs:$a),
(CVT_f64_u32 Int32Regs:$a, CvtRN)>;
def : Pat<(int_nvvm_ui2d_rz Int32Regs:$a),
(CVT_f64_u32 Int32Regs:$a, CvtRZ)>;
def : Pat<(int_nvvm_ui2d_rm Int32Regs:$a),
(CVT_f64_u32 Int32Regs:$a, CvtRM)>;
def : Pat<(int_nvvm_ui2d_rp Int32Regs:$a),
(CVT_f64_u32 Int32Regs:$a, CvtRP)>;
def : Pat<(int_nvvm_f2i_rn_ftz Float32Regs:$a),
(CVT_s32_f32 Float32Regs:$a, CvtRNI_FTZ)>;
def : Pat<(int_nvvm_f2i_rn Float32Regs:$a),
(CVT_s32_f32 Float32Regs:$a, CvtRNI)>;
def : Pat<(int_nvvm_f2i_rz_ftz Float32Regs:$a),
(CVT_s32_f32 Float32Regs:$a, CvtRZI_FTZ)>;
def : Pat<(int_nvvm_f2i_rz Float32Regs:$a),
(CVT_s32_f32 Float32Regs:$a, CvtRZI)>;
def : Pat<(int_nvvm_f2i_rm_ftz Float32Regs:$a),
(CVT_s32_f32 Float32Regs:$a, CvtRMI_FTZ)>;
def : Pat<(int_nvvm_f2i_rm Float32Regs:$a),
(CVT_s32_f32 Float32Regs:$a, CvtRMI)>;
def : Pat<(int_nvvm_f2i_rp_ftz Float32Regs:$a),
(CVT_s32_f32 Float32Regs:$a, CvtRPI_FTZ)>;
def : Pat<(int_nvvm_f2i_rp Float32Regs:$a),
(CVT_s32_f32 Float32Regs:$a, CvtRPI)>;
def : Pat<(int_nvvm_f2ui_rn_ftz Float32Regs:$a),
(CVT_u32_f32 Float32Regs:$a, CvtRNI_FTZ)>;
def : Pat<(int_nvvm_f2ui_rn Float32Regs:$a),
(CVT_u32_f32 Float32Regs:$a, CvtRNI)>;
def : Pat<(int_nvvm_f2ui_rz_ftz Float32Regs:$a),
(CVT_u32_f32 Float32Regs:$a, CvtRZI_FTZ)>;
def : Pat<(int_nvvm_f2ui_rz Float32Regs:$a),
(CVT_u32_f32 Float32Regs:$a, CvtRZI)>;
def : Pat<(int_nvvm_f2ui_rm_ftz Float32Regs:$a),
(CVT_u32_f32 Float32Regs:$a, CvtRMI_FTZ)>;
def : Pat<(int_nvvm_f2ui_rm Float32Regs:$a),
(CVT_u32_f32 Float32Regs:$a, CvtRMI)>;
def : Pat<(int_nvvm_f2ui_rp_ftz Float32Regs:$a),
(CVT_u32_f32 Float32Regs:$a, CvtRPI_FTZ)>;
def : Pat<(int_nvvm_f2ui_rp Float32Regs:$a),
(CVT_u32_f32 Float32Regs:$a, CvtRPI)>;
def : Pat<(int_nvvm_i2f_rn Int32Regs:$a),
(CVT_f32_s32 Int32Regs:$a, CvtRN)>;
def : Pat<(int_nvvm_i2f_rz Int32Regs:$a),
(CVT_f32_s32 Int32Regs:$a, CvtRZ)>;
def : Pat<(int_nvvm_i2f_rm Int32Regs:$a),
(CVT_f32_s32 Int32Regs:$a, CvtRM)>;
def : Pat<(int_nvvm_i2f_rp Int32Regs:$a),
(CVT_f32_s32 Int32Regs:$a, CvtRP)>;
def : Pat<(int_nvvm_ui2f_rn Int32Regs:$a),
(CVT_f32_u32 Int32Regs:$a, CvtRN)>;
def : Pat<(int_nvvm_ui2f_rz Int32Regs:$a),
(CVT_f32_u32 Int32Regs:$a, CvtRZ)>;
def : Pat<(int_nvvm_ui2f_rm Int32Regs:$a),
(CVT_f32_u32 Int32Regs:$a, CvtRM)>;
def : Pat<(int_nvvm_ui2f_rp Int32Regs:$a),
(CVT_f32_u32 Int32Regs:$a, CvtRP)>;
def INT_NVVM_LOHI_I2D : F_MATH_2<"mov.b64 \t$dst, {{$src0, $src1}};",
Float64Regs, Int32Regs, Int32Regs, int_nvvm_lohi_i2d>;
def INT_NVVM_D2I_LO : F_MATH_1<
!strconcat("{{\n\t",
".reg .b32 %temp; \n\t",
"mov.b64 \t{$dst, %temp}, $src0;\n\t",
"}}"),
Int32Regs, Float64Regs, int_nvvm_d2i_lo>;
def INT_NVVM_D2I_HI : F_MATH_1<
!strconcat("{{\n\t",
".reg .b32 %temp; \n\t",
"mov.b64 \t{%temp, $dst}, $src0;\n\t",
"}}"),
Int32Regs, Float64Regs, int_nvvm_d2i_hi>;
def : Pat<(int_nvvm_f2ll_rn_ftz Float32Regs:$a),
(CVT_s64_f32 Float32Regs:$a, CvtRNI_FTZ)>;
def : Pat<(int_nvvm_f2ll_rn Float32Regs:$a),
(CVT_s64_f32 Float32Regs:$a, CvtRNI)>;
def : Pat<(int_nvvm_f2ll_rz_ftz Float32Regs:$a),
(CVT_s64_f32 Float32Regs:$a, CvtRZI_FTZ)>;
def : Pat<(int_nvvm_f2ll_rz Float32Regs:$a),
(CVT_s64_f32 Float32Regs:$a, CvtRZI)>;
def : Pat<(int_nvvm_f2ll_rm_ftz Float32Regs:$a),
(CVT_s64_f32 Float32Regs:$a, CvtRMI_FTZ)>;
def : Pat<(int_nvvm_f2ll_rm Float32Regs:$a),
(CVT_s64_f32 Float32Regs:$a, CvtRMI)>;
def : Pat<(int_nvvm_f2ll_rp_ftz Float32Regs:$a),
(CVT_s64_f32 Float32Regs:$a, CvtRPI_FTZ)>;
def : Pat<(int_nvvm_f2ll_rp Float32Regs:$a),
(CVT_s64_f32 Float32Regs:$a, CvtRPI)>;
def : Pat<(int_nvvm_f2ull_rn_ftz Float32Regs:$a),
(CVT_u64_f32 Float32Regs:$a, CvtRNI_FTZ)>;
def : Pat<(int_nvvm_f2ull_rn Float32Regs:$a),
(CVT_u64_f32 Float32Regs:$a, CvtRNI)>;
def : Pat<(int_nvvm_f2ull_rz_ftz Float32Regs:$a),
(CVT_u64_f32 Float32Regs:$a, CvtRZI_FTZ)>;
def : Pat<(int_nvvm_f2ull_rz Float32Regs:$a),
(CVT_u64_f32 Float32Regs:$a, CvtRZI)>;
def : Pat<(int_nvvm_f2ull_rm_ftz Float32Regs:$a),
(CVT_u64_f32 Float32Regs:$a, CvtRMI_FTZ)>;
def : Pat<(int_nvvm_f2ull_rm Float32Regs:$a),
(CVT_u64_f32 Float32Regs:$a, CvtRMI)>;
def : Pat<(int_nvvm_f2ull_rp_ftz Float32Regs:$a),
(CVT_u64_f32 Float32Regs:$a, CvtRPI_FTZ)>;
def : Pat<(int_nvvm_f2ull_rp Float32Regs:$a),
(CVT_u64_f32 Float32Regs:$a, CvtRPI)>;
def : Pat<(int_nvvm_d2ll_rn Float64Regs:$a),
(CVT_s64_f64 Float64Regs:$a, CvtRNI)>;
def : Pat<(int_nvvm_d2ll_rz Float64Regs:$a),
(CVT_s64_f64 Float64Regs:$a, CvtRZI)>;
def : Pat<(int_nvvm_d2ll_rm Float64Regs:$a),
(CVT_s64_f64 Float64Regs:$a, CvtRMI)>;
def : Pat<(int_nvvm_d2ll_rp Float64Regs:$a),
(CVT_s64_f64 Float64Regs:$a, CvtRPI)>;
def : Pat<(int_nvvm_d2ull_rn Float64Regs:$a),
(CVT_u64_f64 Float64Regs:$a, CvtRNI)>;
def : Pat<(int_nvvm_d2ull_rz Float64Regs:$a),
(CVT_u64_f64 Float64Regs:$a, CvtRZI)>;
def : Pat<(int_nvvm_d2ull_rm Float64Regs:$a),
(CVT_u64_f64 Float64Regs:$a, CvtRMI)>;
def : Pat<(int_nvvm_d2ull_rp Float64Regs:$a),
(CVT_u64_f64 Float64Regs:$a, CvtRPI)>;
def : Pat<(int_nvvm_ll2f_rn Int64Regs:$a),
(CVT_f32_s64 Int64Regs:$a, CvtRN)>;
def : Pat<(int_nvvm_ll2f_rz Int64Regs:$a),
(CVT_f32_s64 Int64Regs:$a, CvtRZ)>;
def : Pat<(int_nvvm_ll2f_rm Int64Regs:$a),
(CVT_f32_s64 Int64Regs:$a, CvtRM)>;
def : Pat<(int_nvvm_ll2f_rp Int64Regs:$a),
(CVT_f32_s64 Int64Regs:$a, CvtRP)>;
def : Pat<(int_nvvm_ull2f_rn Int64Regs:$a),
(CVT_f32_u64 Int64Regs:$a, CvtRN)>;
def : Pat<(int_nvvm_ull2f_rz Int64Regs:$a),
(CVT_f32_u64 Int64Regs:$a, CvtRZ)>;
def : Pat<(int_nvvm_ull2f_rm Int64Regs:$a),
(CVT_f32_u64 Int64Regs:$a, CvtRM)>;
def : Pat<(int_nvvm_ull2f_rp Int64Regs:$a),
(CVT_f32_u64 Int64Regs:$a, CvtRP)>;
def : Pat<(int_nvvm_ll2d_rn Int64Regs:$a),
(CVT_f64_s64 Int64Regs:$a, CvtRN)>;
def : Pat<(int_nvvm_ll2d_rz Int64Regs:$a),
(CVT_f64_s64 Int64Regs:$a, CvtRZ)>;
def : Pat<(int_nvvm_ll2d_rm Int64Regs:$a),
(CVT_f64_s64 Int64Regs:$a, CvtRM)>;
def : Pat<(int_nvvm_ll2d_rp Int64Regs:$a),
(CVT_f64_s64 Int64Regs:$a, CvtRP)>;
def : Pat<(int_nvvm_ull2d_rn Int64Regs:$a),
(CVT_f64_u64 Int64Regs:$a, CvtRN)>;
def : Pat<(int_nvvm_ull2d_rz Int64Regs:$a),
(CVT_f64_u64 Int64Regs:$a, CvtRZ)>;
def : Pat<(int_nvvm_ull2d_rm Int64Regs:$a),
(CVT_f64_u64 Int64Regs:$a, CvtRM)>;
def : Pat<(int_nvvm_ull2d_rp Int64Regs:$a),
(CVT_f64_u64 Int64Regs:$a, CvtRP)>;
def : Pat<(int_nvvm_f2h_rn_ftz Float32Regs:$a),
(BITCONVERT_16_F2I (CVT_f16_f32 Float32Regs:$a, CvtRN_FTZ))>;
def : Pat<(int_nvvm_f2h_rn Float32Regs:$a),
(BITCONVERT_16_F2I (CVT_f16_f32 Float32Regs:$a, CvtRN))>;
//
// Bitcast
//
def INT_NVVM_BITCAST_F2I : F_MATH_1<"mov.b32 \t$dst, $src0;", Int32Regs,
Float32Regs, int_nvvm_bitcast_f2i>;
def INT_NVVM_BITCAST_I2F : F_MATH_1<"mov.b32 \t$dst, $src0;", Float32Regs,
Int32Regs, int_nvvm_bitcast_i2f>;
def INT_NVVM_BITCAST_LL2D : F_MATH_1<"mov.b64 \t$dst, $src0;", Float64Regs,
Int64Regs, int_nvvm_bitcast_ll2d>;
def INT_NVVM_BITCAST_D2LL : F_MATH_1<"mov.b64 \t$dst, $src0;", Int64Regs,
Float64Regs, int_nvvm_bitcast_d2ll>;
//
// FNS
//
class INT_FNS_MBO<dag ins, dag Operands>
: NVPTXInst<(outs Int32Regs:$dst), ins,
"fns.b32 \t$dst, $mask, $base, $offset;",
[(set Int32Regs:$dst, Operands )]>,
Requires<[hasPTX60, hasSM30]>;
def INT_FNS_rrr : INT_FNS_MBO<(ins Int32Regs:$mask, Int32Regs:$base, Int32Regs:$offset),
(int_nvvm_fns Int32Regs:$mask, Int32Regs:$base, Int32Regs:$offset)>;
def INT_FNS_rri : INT_FNS_MBO<(ins Int32Regs:$mask, Int32Regs:$base, i32imm:$offset),
(int_nvvm_fns Int32Regs:$mask, Int32Regs:$base, imm:$offset)>;
def INT_FNS_rir : INT_FNS_MBO<(ins Int32Regs:$mask, i32imm:$base, Int32Regs:$offset),
(int_nvvm_fns Int32Regs:$mask, imm:$base, Int32Regs:$offset)>;
def INT_FNS_rii : INT_FNS_MBO<(ins Int32Regs:$mask, i32imm:$base, i32imm:$offset),
(int_nvvm_fns Int32Regs:$mask, imm:$base, imm:$offset)>;
def INT_FNS_irr : INT_FNS_MBO<(ins i32imm:$mask, Int32Regs:$base, Int32Regs:$offset),
(int_nvvm_fns imm:$mask, Int32Regs:$base, Int32Regs:$offset)>;
def INT_FNS_iri : INT_FNS_MBO<(ins i32imm:$mask, Int32Regs:$base, i32imm:$offset),
(int_nvvm_fns imm:$mask, Int32Regs:$base, imm:$offset)>;
def INT_FNS_iir : INT_FNS_MBO<(ins i32imm:$mask, i32imm:$base, Int32Regs:$offset),
(int_nvvm_fns imm:$mask, imm:$base, Int32Regs:$offset)>;
def INT_FNS_iii : INT_FNS_MBO<(ins i32imm:$mask, i32imm:$base, i32imm:$offset),
(int_nvvm_fns imm:$mask, imm:$base, imm:$offset)>;
//-----------------------------------
// Atomic Functions
//-----------------------------------
class ATOMIC_GLOBAL_CHK <dag ops, dag frag>
: PatFrag<ops, frag, AS_match.global>;
class ATOMIC_SHARED_CHK <dag ops, dag frag>
: PatFrag<ops, frag, AS_match.shared>;
class ATOMIC_GENERIC_CHK <dag ops, dag frag>
: PatFrag<ops, frag, AS_match.generic>;
multiclass F_ATOMIC_2_imp<NVPTXRegClass ptrclass, NVPTXRegClass regclass,
string SpaceStr, string TypeStr, string OpcStr, PatFrag IntOp,
Operand IMMType, SDNode IMM, list<Predicate> Pred> {
def reg : NVPTXInst<(outs regclass:$dst), (ins ptrclass:$addr, regclass:$b),
!strconcat("atom", SpaceStr, OpcStr, TypeStr, " \t$dst, [$addr], $b;"),
[(set regclass:$dst, (IntOp ptrclass:$addr, regclass:$b))]>,
Requires<Pred>;
def imm : NVPTXInst<(outs regclass:$dst), (ins ptrclass:$addr, IMMType:$b),
!strconcat("atom", SpaceStr, OpcStr, TypeStr, " \t$dst, [$addr], $b;", ""),
[(set regclass:$dst, (IntOp ptrclass:$addr, IMM:$b))]>,
Requires<Pred>;
}
multiclass F_ATOMIC_2<NVPTXRegClass regclass, string SpaceStr, string TypeStr,
string OpcStr, PatFrag IntOp, Operand IMMType, SDNode IMM,
list<Predicate> Pred = []> {
defm p32 : F_ATOMIC_2_imp<Int32Regs, regclass, SpaceStr, TypeStr, OpcStr,
IntOp, IMMType, IMM, Pred>;
defm p64 : F_ATOMIC_2_imp<Int64Regs, regclass, SpaceStr, TypeStr, OpcStr,
IntOp, IMMType, IMM, Pred>;
}
// has 2 operands, neg the second one
multiclass F_ATOMIC_2_NEG_imp<NVPTXRegClass ptrclass, NVPTXRegClass regclass,
string SpaceStr, string TypeStr, string OpcStr, PatFrag IntOp,
Operand IMMType, list<Predicate> Pred> {
def reg : NVPTXInst<(outs regclass:$dst), (ins ptrclass:$addr, regclass:$b),
!strconcat(
"{{ \n\t",
".reg \t.s", TypeStr, " temp; \n\t",
"neg.s", TypeStr, " \ttemp, $b; \n\t",
"atom", SpaceStr, OpcStr, ".u", TypeStr, " \t$dst, [$addr], temp; \n\t",
"}}"),
[(set regclass:$dst, (IntOp ptrclass:$addr, regclass:$b))]>,
Requires<Pred>;
}
multiclass F_ATOMIC_2_NEG<NVPTXRegClass regclass, string SpaceStr,
string TypeStr, string OpcStr, PatFrag IntOp, Operand IMMType,
list<Predicate> Pred = []> {
defm p32: F_ATOMIC_2_NEG_imp<Int32Regs, regclass, SpaceStr, TypeStr, OpcStr,
IntOp, IMMType, Pred> ;
defm p64: F_ATOMIC_2_NEG_imp<Int64Regs, regclass, SpaceStr, TypeStr, OpcStr,
IntOp, IMMType, Pred> ;
}
// has 3 operands
multiclass F_ATOMIC_3_imp<NVPTXRegClass ptrclass, NVPTXRegClass regclass,
string SpaceStr, string TypeStr, string OpcStr, PatFrag IntOp,
Operand IMMType, list<Predicate> Pred> {
def reg : NVPTXInst<(outs regclass:$dst),
(ins ptrclass:$addr, regclass:$b, regclass:$c),
!strconcat("atom", SpaceStr, OpcStr, TypeStr, " \t$dst, [$addr], $b, $c;"),
[(set regclass:$dst, (IntOp ptrclass:$addr, regclass:$b, regclass:$c))]>,
Requires<Pred>;
def imm1 : NVPTXInst<(outs regclass:$dst),
(ins ptrclass:$addr, IMMType:$b, regclass:$c),
!strconcat("atom", SpaceStr, OpcStr, TypeStr, " \t$dst, [$addr], $b, $c;"),
[(set regclass:$dst, (IntOp ptrclass:$addr, imm:$b, regclass:$c))]>,
Requires<Pred>;
def imm2 : NVPTXInst<(outs regclass:$dst),
(ins ptrclass:$addr, regclass:$b, IMMType:$c),
!strconcat("atom", SpaceStr, OpcStr, TypeStr, " \t$dst, [$addr], $b, $c;", ""),
[(set regclass:$dst, (IntOp ptrclass:$addr, regclass:$b, imm:$c))]>,
Requires<Pred>;
def imm3 : NVPTXInst<(outs regclass:$dst),
(ins ptrclass:$addr, IMMType:$b, IMMType:$c),
!strconcat("atom", SpaceStr, OpcStr, TypeStr, " \t$dst, [$addr], $b, $c;"),
[(set regclass:$dst, (IntOp ptrclass:$addr, imm:$b, imm:$c))]>,
Requires<Pred>;
}
multiclass F_ATOMIC_3<NVPTXRegClass regclass, string SpaceStr, string TypeStr,
string OpcStr, PatFrag IntOp, Operand IMMType, list<Predicate> Pred = []> {
defm p32 : F_ATOMIC_3_imp<Int32Regs, regclass, SpaceStr, TypeStr, OpcStr,
IntOp, IMMType, Pred>;
defm p64 : F_ATOMIC_3_imp<Int64Regs, regclass, SpaceStr, TypeStr, OpcStr,
IntOp, IMMType, Pred>;
}
// atom_add
def atomic_load_add_32_g: ATOMIC_GLOBAL_CHK<(ops node:$a, node:$b),
(atomic_load_add_32 node:$a, node:$b)>;
def atomic_load_add_32_s: ATOMIC_SHARED_CHK<(ops node:$a, node:$b),
(atomic_load_add_32 node:$a, node:$b)>;
def atomic_load_add_32_gen: ATOMIC_GENERIC_CHK<(ops node:$a, node:$b),
(atomic_load_add_32 node:$a, node:$b)>;
def atomic_load_add_64_g: ATOMIC_GLOBAL_CHK<(ops node:$a, node:$b),
(atomic_load_add_64 node:$a, node:$b)>;
def atomic_load_add_64_s: ATOMIC_SHARED_CHK<(ops node:$a, node:$b),
(atomic_load_add_64 node:$a, node:$b)>;
def atomic_load_add_64_gen: ATOMIC_GENERIC_CHK<(ops node:$a, node:$b),
(atomic_load_add_64 node:$a, node:$b)>;
def atomic_load_add_f32_g: ATOMIC_GLOBAL_CHK<(ops node:$a, node:$b),
(int_nvvm_atomic_load_add_f32 node:$a, node:$b)>;
def atomic_load_add_f32_s: ATOMIC_SHARED_CHK<(ops node:$a, node:$b),
(int_nvvm_atomic_load_add_f32 node:$a, node:$b)>;
def atomic_load_add_f32_gen: ATOMIC_GENERIC_CHK<(ops node:$a, node:$b),
(int_nvvm_atomic_load_add_f32 node:$a, node:$b)>;
def atomic_load_add_f64_g: ATOMIC_GLOBAL_CHK<(ops node:$a, node:$b),
(int_nvvm_atomic_load_add_f64 node:$a, node:$b)>;
def atomic_load_add_f64_s: ATOMIC_SHARED_CHK<(ops node:$a, node:$b),
(int_nvvm_atomic_load_add_f64 node:$a, node:$b)>;
def atomic_load_add_f64_gen: ATOMIC_GENERIC_CHK<(ops node:$a, node:$b),
(int_nvvm_atomic_load_add_f64 node:$a, node:$b)>;
defm INT_PTX_ATOM_ADD_G_32 : F_ATOMIC_2<Int32Regs, ".global", ".u32", ".add",
atomic_load_add_32_g, i32imm, imm>;
defm INT_PTX_ATOM_ADD_S_32 : F_ATOMIC_2<Int32Regs, ".shared", ".u32", ".add",
atomic_load_add_32_s, i32imm, imm>;
defm INT_PTX_ATOM_ADD_GEN_32 : F_ATOMIC_2<Int32Regs, "", ".u32", ".add",
atomic_load_add_32_gen, i32imm, imm>;
defm INT_PTX_ATOM_ADD_GEN_32_USE_G : F_ATOMIC_2<Int32Regs, ".global", ".u32",
".add", atomic_load_add_32_gen, i32imm, imm>;
defm INT_PTX_ATOM_ADD_G_64 : F_ATOMIC_2<Int64Regs, ".global", ".u64", ".add",
atomic_load_add_64_g, i64imm, imm>;
defm INT_PTX_ATOM_ADD_S_64 : F_ATOMIC_2<Int64Regs, ".shared", ".u64", ".add",
atomic_load_add_64_s, i64imm, imm>;
defm INT_PTX_ATOM_ADD_GEN_64 : F_ATOMIC_2<Int64Regs, "", ".u64", ".add",
atomic_load_add_64_gen, i64imm, imm>;
defm INT_PTX_ATOM_ADD_GEN_64_USE_G : F_ATOMIC_2<Int64Regs, ".global", ".u64",
".add", atomic_load_add_64_gen, i64imm, imm>;
defm INT_PTX_ATOM_ADD_G_F32 : F_ATOMIC_2<Float32Regs, ".global", ".f32", ".add",
atomic_load_add_f32_g, f32imm, fpimm>;
defm INT_PTX_ATOM_ADD_S_F32 : F_ATOMIC_2<Float32Regs, ".shared", ".f32", ".add",
atomic_load_add_f32_s, f32imm, fpimm>;
defm INT_PTX_ATOM_ADD_GEN_F32 : F_ATOMIC_2<Float32Regs, "", ".f32", ".add",
atomic_load_add_f32_gen, f32imm, fpimm>;
defm INT_PTX_ATOM_ADD_G_F64 : F_ATOMIC_2<Float64Regs, ".global", ".f64", ".add",
atomic_load_add_f64_g, f64imm, fpimm, [hasAtomAddF64]>;
defm INT_PTX_ATOM_ADD_S_F64 : F_ATOMIC_2<Float64Regs, ".shared", ".f64", ".add",
atomic_load_add_f64_s, f64imm, fpimm, [hasAtomAddF64]>;
defm INT_PTX_ATOM_ADD_GEN_F64 : F_ATOMIC_2<Float64Regs, "", ".f64", ".add",
atomic_load_add_f64_gen, f64imm, fpimm, [hasAtomAddF64]>;
// atom_sub
def atomic_load_sub_32_g: ATOMIC_GLOBAL_CHK<(ops node:$a, node:$b),
(atomic_load_sub_32 node:$a, node:$b)>;
def atomic_load_sub_32_s: ATOMIC_SHARED_CHK<(ops node:$a, node:$b),
(atomic_load_sub_32 node:$a, node:$b)>;
def atomic_load_sub_32_gen: ATOMIC_GENERIC_CHK<(ops node:$a, node:$b),
(atomic_load_sub_32 node:$a, node:$b)>;
def atomic_load_sub_64_g: ATOMIC_GLOBAL_CHK<(ops node:$a, node:$b),
(atomic_load_sub_64 node:$a, node:$b)>;
def atomic_load_sub_64_s: ATOMIC_SHARED_CHK<(ops node:$a, node:$b),
(atomic_load_sub_64 node:$a, node:$b)>;
def atomic_load_sub_64_gen: ATOMIC_GENERIC_CHK<(ops node:$a, node:$b),
(atomic_load_sub_64 node:$a, node:$b)>;
defm INT_PTX_ATOM_SUB_G_32 : F_ATOMIC_2_NEG<Int32Regs, ".global", "32", ".add",
atomic_load_sub_32_g, i32imm>;
defm INT_PTX_ATOM_SUB_G_64 : F_ATOMIC_2_NEG<Int64Regs, ".global", "64", ".add",
atomic_load_sub_64_g, i64imm>;
defm INT_PTX_ATOM_SUB_GEN_32 : F_ATOMIC_2_NEG<Int32Regs, "", "32", ".add",
atomic_load_sub_32_gen, i32imm>;
defm INT_PTX_ATOM_SUB_GEN_32_USE_G : F_ATOMIC_2_NEG<Int32Regs, ".global", "32",
".add", atomic_load_sub_32_gen, i32imm>;
defm INT_PTX_ATOM_SUB_S_32 : F_ATOMIC_2_NEG<Int32Regs, ".shared", "32", ".add",
atomic_load_sub_32_s, i32imm>;
defm INT_PTX_ATOM_SUB_S_64 : F_ATOMIC_2_NEG<Int64Regs, ".shared", "64", ".add",
atomic_load_sub_64_s, i64imm>;
defm INT_PTX_ATOM_SUB_GEN_64 : F_ATOMIC_2_NEG<Int64Regs, "", "64", ".add",
atomic_load_sub_64_gen, i64imm>;
defm INT_PTX_ATOM_SUB_GEN_64_USE_G : F_ATOMIC_2_NEG<Int64Regs, ".global", "64",
".add", atomic_load_sub_64_gen, i64imm>;
// atom_swap
def atomic_swap_32_g: ATOMIC_GLOBAL_CHK<(ops node:$a, node:$b),
(atomic_swap_32 node:$a, node:$b)>;
def atomic_swap_32_s: ATOMIC_SHARED_CHK<(ops node:$a, node:$b),
(atomic_swap_32 node:$a, node:$b)>;
def atomic_swap_32_gen: ATOMIC_GENERIC_CHK<(ops node:$a, node:$b),
(atomic_swap_32 node:$a, node:$b)>;
def atomic_swap_64_g: ATOMIC_GLOBAL_CHK<(ops node:$a, node:$b),
(atomic_swap_64 node:$a, node:$b)>;
def atomic_swap_64_s: ATOMIC_SHARED_CHK<(ops node:$a, node:$b),
(atomic_swap_64 node:$a, node:$b)>;
def atomic_swap_64_gen: ATOMIC_GENERIC_CHK<(ops node:$a, node:$b),
(atomic_swap_64 node:$a, node:$b)>;
defm INT_PTX_ATOM_SWAP_G_32 : F_ATOMIC_2<Int32Regs, ".global", ".b32", ".exch",
atomic_swap_32_g, i32imm, imm>;
defm INT_PTX_ATOM_SWAP_S_32 : F_ATOMIC_2<Int32Regs, ".shared", ".b32", ".exch",
atomic_swap_32_s, i32imm, imm>;
defm INT_PTX_ATOM_SWAP_GEN_32 : F_ATOMIC_2<Int32Regs, "", ".b32", ".exch",
atomic_swap_32_gen, i32imm, imm>;
defm INT_PTX_ATOM_SWAP_GEN_32_USE_G : F_ATOMIC_2<Int32Regs, ".global", ".b32",
".exch", atomic_swap_32_gen, i32imm, imm>;
defm INT_PTX_ATOM_SWAP_G_64 : F_ATOMIC_2<Int64Regs, ".global", ".b64", ".exch",
atomic_swap_64_g, i64imm, imm>;
defm INT_PTX_ATOM_SWAP_S_64 : F_ATOMIC_2<Int64Regs, ".shared", ".b64", ".exch",
atomic_swap_64_s, i64imm, imm>;
defm INT_PTX_ATOM_SWAP_GEN_64 : F_ATOMIC_2<Int64Regs, "", ".b64", ".exch",
atomic_swap_64_gen, i64imm, imm>;
defm INT_PTX_ATOM_SWAP_GEN_64_USE_G : F_ATOMIC_2<Int64Regs, ".global", ".b64",
".exch", atomic_swap_64_gen, i64imm, imm>;
// atom_max
def atomic_load_max_32_g: ATOMIC_GLOBAL_CHK<(ops node:$a, node:$b)
, (atomic_load_max_32 node:$a, node:$b)>;
def atomic_load_max_32_s: ATOMIC_SHARED_CHK<(ops node:$a, node:$b),
(atomic_load_max_32 node:$a, node:$b)>;
def atomic_load_max_32_gen: ATOMIC_GENERIC_CHK<(ops node:$a, node:$b),
(atomic_load_max_32 node:$a, node:$b)>;
def atomic_load_max_64_g: ATOMIC_GLOBAL_CHK<(ops node:$a, node:$b)
, (atomic_load_max_64 node:$a, node:$b)>;
def atomic_load_max_64_s: ATOMIC_SHARED_CHK<(ops node:$a, node:$b),
(atomic_load_max_64 node:$a, node:$b)>;
def atomic_load_max_64_gen: ATOMIC_GENERIC_CHK<(ops node:$a, node:$b),
(atomic_load_max_64 node:$a, node:$b)>;
def atomic_load_umax_32_g: ATOMIC_GLOBAL_CHK<(ops node:$a, node:$b),
(atomic_load_umax_32 node:$a, node:$b)>;
def atomic_load_umax_32_s: ATOMIC_SHARED_CHK<(ops node:$a, node:$b),
(atomic_load_umax_32 node:$a, node:$b)>;
def atomic_load_umax_32_gen: ATOMIC_GENERIC_CHK<(ops node:$a, node:$b),
(atomic_load_umax_32 node:$a, node:$b)>;
def atomic_load_umax_64_g: ATOMIC_GLOBAL_CHK<(ops node:$a, node:$b),
(atomic_load_umax_64 node:$a, node:$b)>;
def atomic_load_umax_64_s: ATOMIC_SHARED_CHK<(ops node:$a, node:$b),
(atomic_load_umax_64 node:$a, node:$b)>;
def atomic_load_umax_64_gen: ATOMIC_GENERIC_CHK<(ops node:$a, node:$b),
(atomic_load_umax_64 node:$a, node:$b)>;
defm INT_PTX_ATOM_LOAD_MAX_G_32 : F_ATOMIC_2<Int32Regs, ".global", ".s32",
".max", atomic_load_max_32_g, i32imm, imm>;
defm INT_PTX_ATOM_LOAD_MAX_S_32 : F_ATOMIC_2<Int32Regs, ".shared", ".s32",
".max", atomic_load_max_32_s, i32imm, imm>;
defm INT_PTX_ATOM_LOAD_MAX_GEN_32 : F_ATOMIC_2<Int32Regs, "", ".s32", ".max",
atomic_load_max_32_gen, i32imm, imm>;
defm INT_PTX_ATOM_LOAD_MAX_GEN_32_USE_G : F_ATOMIC_2<Int32Regs, ".global",
".s32", ".max", atomic_load_max_32_gen, i32imm, imm>;
defm INT_PTX_ATOM_LOAD_MAX_G_64 : F_ATOMIC_2<Int64Regs, ".global", ".s64",
".max", atomic_load_max_64_g, i64imm, imm>;
defm INT_PTX_ATOM_LOAD_MAX_S_64 : F_ATOMIC_2<Int64Regs, ".shared", ".s64",
".max", atomic_load_max_64_s, i64imm, imm>;
defm INT_PTX_ATOM_LOAD_MAX_GEN_64 : F_ATOMIC_2<Int64Regs, "", ".s64", ".max",
atomic_load_max_64_gen, i64imm, imm>;
defm INT_PTX_ATOM_LOAD_MAX_GEN_64_USE_G : F_ATOMIC_2<Int64Regs, ".global",
".s64", ".max", atomic_load_max_64_gen, i64imm, imm>;
defm INT_PTX_ATOM_LOAD_UMAX_G_32 : F_ATOMIC_2<Int32Regs, ".global", ".u32",
".max", atomic_load_umax_32_g, i32imm, imm>;
defm INT_PTX_ATOM_LOAD_UMAX_S_32 : F_ATOMIC_2<Int32Regs, ".shared", ".u32",
".max", atomic_load_umax_32_s, i32imm, imm>;
defm INT_PTX_ATOM_LOAD_UMAX_GEN_32 : F_ATOMIC_2<Int32Regs, "", ".u32", ".max",
atomic_load_umax_32_gen, i32imm, imm>;
defm INT_PTX_ATOM_LOAD_UMAX_GEN_32_USE_G : F_ATOMIC_2<Int32Regs, ".global",
".u32", ".max", atomic_load_umax_32_gen, i32imm, imm>;
defm INT_PTX_ATOM_LOAD_UMAX_G_64 : F_ATOMIC_2<Int64Regs, ".global", ".u64",
".max", atomic_load_umax_64_g, i64imm, imm>;
defm INT_PTX_ATOM_LOAD_UMAX_S_64 : F_ATOMIC_2<Int64Regs, ".shared", ".u64",
".max", atomic_load_umax_64_s, i64imm, imm>;
defm INT_PTX_ATOM_LOAD_UMAX_GEN_64 : F_ATOMIC_2<Int64Regs, "", ".u64", ".max",
atomic_load_umax_64_gen, i64imm, imm>;
defm INT_PTX_ATOM_LOAD_UMAX_GEN_64_USE_G : F_ATOMIC_2<Int64Regs, ".global",
".u64", ".max", atomic_load_umax_64_gen, i64imm, imm>;
// atom_min
def atomic_load_min_32_g: ATOMIC_GLOBAL_CHK<(ops node:$a, node:$b),
(atomic_load_min_32 node:$a, node:$b)>;
def atomic_load_min_32_s: ATOMIC_SHARED_CHK<(ops node:$a, node:$b),
(atomic_load_min_32 node:$a, node:$b)>;
def atomic_load_min_32_gen: ATOMIC_GENERIC_CHK<(ops node:$a, node:$b),
(atomic_load_min_32 node:$a, node:$b)>;
def atomic_load_min_64_g: ATOMIC_GLOBAL_CHK<(ops node:$a, node:$b),
(atomic_load_min_64 node:$a, node:$b)>;
def atomic_load_min_64_s: ATOMIC_SHARED_CHK<(ops node:$a, node:$b),
(atomic_load_min_64 node:$a, node:$b)>;
def atomic_load_min_64_gen: ATOMIC_GENERIC_CHK<(ops node:$a, node:$b),
(atomic_load_min_64 node:$a, node:$b)>;
def atomic_load_umin_32_g: ATOMIC_GLOBAL_CHK<(ops node:$a, node:$b),
(atomic_load_umin_32 node:$a, node:$b)>;
def atomic_load_umin_32_s: ATOMIC_SHARED_CHK<(ops node:$a, node:$b),
(atomic_load_umin_32 node:$a, node:$b)>;
def atomic_load_umin_32_gen: ATOMIC_GENERIC_CHK<(ops node:$a, node:$b),
(atomic_load_umin_32 node:$a, node:$b)>;
def atomic_load_umin_64_g: ATOMIC_GLOBAL_CHK<(ops node:$a, node:$b),
(atomic_load_umin_64 node:$a, node:$b)>;
def atomic_load_umin_64_s: ATOMIC_SHARED_CHK<(ops node:$a, node:$b),
(atomic_load_umin_64 node:$a, node:$b)>;
def atomic_load_umin_64_gen: ATOMIC_GENERIC_CHK<(ops node:$a, node:$b),
(atomic_load_umin_64 node:$a, node:$b)>;
defm INT_PTX_ATOM_LOAD_MIN_G_32 : F_ATOMIC_2<Int32Regs, ".global", ".s32",
".min", atomic_load_min_32_g, i32imm, imm>;
defm INT_PTX_ATOM_LOAD_MIN_S_32 : F_ATOMIC_2<Int32Regs, ".shared", ".s32",
".min", atomic_load_min_32_s, i32imm, imm>;
defm INT_PTX_ATOM_LOAD_MIN_GEN_32 : F_ATOMIC_2<Int32Regs, "", ".s32", ".min",
atomic_load_min_32_gen, i32imm, imm>;
defm INT_PTX_ATOM_LOAD_MIN_GEN_32_USE_G : F_ATOMIC_2<Int32Regs, ".global",
".s32", ".min", atomic_load_min_32_gen, i32imm, imm>;
defm INT_PTX_ATOM_LOAD_MIN_G_64 : F_ATOMIC_2<Int64Regs, ".global", ".s64",
".min", atomic_load_min_64_g, i64imm, imm>;
defm INT_PTX_ATOM_LOAD_MIN_S_64 : F_ATOMIC_2<Int64Regs, ".shared", ".s64",
".min", atomic_load_min_64_s, i64imm, imm>;
defm INT_PTX_ATOM_LOAD_MIN_GEN_64 : F_ATOMIC_2<Int64Regs, "", ".s64", ".min",
atomic_load_min_64_gen, i64imm, imm>;
defm INT_PTX_ATOM_LOAD_MIN_GEN_64_USE_G : F_ATOMIC_2<Int64Regs, ".global",
".s64", ".min", atomic_load_min_64_gen, i64imm, imm>;
defm INT_PTX_ATOM_LOAD_UMIN_G_32 : F_ATOMIC_2<Int32Regs, ".global", ".u32",
".min", atomic_load_umin_32_g, i32imm, imm>;
defm INT_PTX_ATOM_LOAD_UMIN_S_32 : F_ATOMIC_2<Int32Regs, ".shared", ".u32",
".min", atomic_load_umin_32_s, i32imm, imm>;
defm INT_PTX_ATOM_LOAD_UMIN_GEN_32 : F_ATOMIC_2<Int32Regs, "", ".u32", ".min",
atomic_load_umin_32_gen, i32imm, imm>;
defm INT_PTX_ATOM_LOAD_UMIN_GEN_32_USE_G : F_ATOMIC_2<Int32Regs, ".global",
".u32", ".min", atomic_load_umin_32_gen, i32imm, imm>;
defm INT_PTX_ATOM_LOAD_UMIN_G_64 : F_ATOMIC_2<Int64Regs, ".global", ".u64",
".min", atomic_load_umin_64_g, i64imm, imm>;
defm INT_PTX_ATOM_LOAD_UMIN_S_64 : F_ATOMIC_2<Int64Regs, ".shared", ".u64",
".min", atomic_load_umin_64_s, i64imm, imm>;
defm INT_PTX_ATOM_LOAD_UMIN_GEN_64 : F_ATOMIC_2<Int64Regs, "", ".u64", ".min",
atomic_load_umin_64_gen, i64imm, imm>;
defm INT_PTX_ATOM_LOAD_UMIN_GEN_64_USE_G : F_ATOMIC_2<Int64Regs, ".global",
".u64", ".min", atomic_load_umin_64_gen, i64imm, imm>;
// atom_inc atom_dec
def atomic_load_inc_32_g: ATOMIC_GLOBAL_CHK<(ops node:$a, node:$b),
(int_nvvm_atomic_load_inc_32 node:$a, node:$b)>;
def atomic_load_inc_32_s: ATOMIC_SHARED_CHK<(ops node:$a, node:$b),
(int_nvvm_atomic_load_inc_32 node:$a, node:$b)>;
def atomic_load_inc_32_gen: ATOMIC_GENERIC_CHK<(ops node:$a, node:$b),
(int_nvvm_atomic_load_inc_32 node:$a, node:$b)>;
def atomic_load_dec_32_g: ATOMIC_GLOBAL_CHK<(ops node:$a, node:$b),
(int_nvvm_atomic_load_dec_32 node:$a, node:$b)>;
def atomic_load_dec_32_s: ATOMIC_SHARED_CHK<(ops node:$a, node:$b),
(int_nvvm_atomic_load_dec_32 node:$a, node:$b)>;
def atomic_load_dec_32_gen: ATOMIC_GENERIC_CHK<(ops node:$a, node:$b),
(int_nvvm_atomic_load_dec_32 node:$a, node:$b)>;
defm INT_PTX_ATOM_INC_G_32 : F_ATOMIC_2<Int32Regs, ".global", ".u32", ".inc",
atomic_load_inc_32_g, i32imm, imm>;
defm INT_PTX_ATOM_INC_S_32 : F_ATOMIC_2<Int32Regs, ".shared", ".u32", ".inc",
atomic_load_inc_32_s, i32imm, imm>;
defm INT_PTX_ATOM_INC_GEN_32 : F_ATOMIC_2<Int32Regs, "", ".u32", ".inc",
atomic_load_inc_32_gen, i32imm, imm>;
defm INT_PTX_ATOM_INC_GEN_32_USE_G : F_ATOMIC_2<Int32Regs, ".global", ".u32",
".inc", atomic_load_inc_32_gen, i32imm, imm>;
defm INT_PTX_ATOM_DEC_G_32 : F_ATOMIC_2<Int32Regs, ".global", ".u32", ".dec",
atomic_load_dec_32_g, i32imm, imm>;
defm INT_PTX_ATOM_DEC_S_32 : F_ATOMIC_2<Int32Regs, ".shared", ".u32", ".dec",
atomic_load_dec_32_s, i32imm, imm>;
defm INT_PTX_ATOM_DEC_GEN_32 : F_ATOMIC_2<Int32Regs, "", ".u32", ".dec",
atomic_load_dec_32_gen, i32imm, imm>;
defm INT_PTX_ATOM_DEC_GEN_32_USE_G : F_ATOMIC_2<Int32Regs, ".global", ".u32",
".dec", atomic_load_dec_32_gen, i32imm, imm>;
// atom_and
def atomic_load_and_32_g: ATOMIC_GLOBAL_CHK<(ops node:$a, node:$b),
(atomic_load_and_32 node:$a, node:$b)>;
def atomic_load_and_32_s: ATOMIC_SHARED_CHK<(ops node:$a, node:$b),
(atomic_load_and_32 node:$a, node:$b)>;
def atomic_load_and_32_gen: ATOMIC_GENERIC_CHK<(ops node:$a, node:$b),
(atomic_load_and_32 node:$a, node:$b)>;
def atomic_load_and_64_g: ATOMIC_GLOBAL_CHK<(ops node:$a, node:$b),
(atomic_load_and_64 node:$a, node:$b)>;
def atomic_load_and_64_s: ATOMIC_SHARED_CHK<(ops node:$a, node:$b),
(atomic_load_and_64 node:$a, node:$b)>;
def atomic_load_and_64_gen: ATOMIC_GENERIC_CHK<(ops node:$a, node:$b),
(atomic_load_and_64 node:$a, node:$b)>;
defm INT_PTX_ATOM_AND_G_32 : F_ATOMIC_2<Int32Regs, ".global", ".b32", ".and",
atomic_load_and_32_g, i32imm, imm>;
defm INT_PTX_ATOM_AND_S_32 : F_ATOMIC_2<Int32Regs, ".shared", ".b32", ".and",
atomic_load_and_32_s, i32imm, imm>;
defm INT_PTX_ATOM_AND_GEN_32 : F_ATOMIC_2<Int32Regs, "", ".b32", ".and",
atomic_load_and_32_gen, i32imm, imm>;
defm INT_PTX_ATOM_AND_GEN_32_USE_G : F_ATOMIC_2<Int32Regs, ".global", ".b32",
".and", atomic_load_and_32_gen, i32imm, imm>;
defm INT_PTX_ATOM_AND_G_64 : F_ATOMIC_2<Int64Regs, ".global", ".b64", ".and",
atomic_load_and_64_g, i64imm, imm>;
defm INT_PTX_ATOM_AND_S_64 : F_ATOMIC_2<Int64Regs, ".shared", ".b64", ".and",
atomic_load_and_64_s, i64imm, imm>;
defm INT_PTX_ATOM_AND_GEN_64 : F_ATOMIC_2<Int64Regs, "", ".b64", ".and",
atomic_load_and_64_gen, i64imm, imm>;
defm INT_PTX_ATOM_AND_GEN_64_USE_G : F_ATOMIC_2<Int64Regs, ".global", ".b64",
".and", atomic_load_and_64_gen, i64imm, imm>;
// atom_or
def atomic_load_or_32_g: ATOMIC_GLOBAL_CHK<(ops node:$a, node:$b),
(atomic_load_or_32 node:$a, node:$b)>;
def atomic_load_or_32_s: ATOMIC_SHARED_CHK<(ops node:$a, node:$b),
(atomic_load_or_32 node:$a, node:$b)>;
def atomic_load_or_32_gen: ATOMIC_GENERIC_CHK<(ops node:$a, node:$b),
(atomic_load_or_32 node:$a, node:$b)>;
def atomic_load_or_64_g: ATOMIC_GLOBAL_CHK<(ops node:$a, node:$b),
(atomic_load_or_64 node:$a, node:$b)>;
def atomic_load_or_64_s: ATOMIC_SHARED_CHK<(ops node:$a, node:$b),
(atomic_load_or_64 node:$a, node:$b)>;
def atomic_load_or_64_gen: ATOMIC_GENERIC_CHK<(ops node:$a, node:$b),
(atomic_load_or_64 node:$a, node:$b)>;
defm INT_PTX_ATOM_OR_G_32 : F_ATOMIC_2<Int32Regs, ".global", ".b32", ".or",
atomic_load_or_32_g, i32imm, imm>;
defm INT_PTX_ATOM_OR_GEN_32 : F_ATOMIC_2<Int32Regs, "", ".b32", ".or",
atomic_load_or_32_gen, i32imm, imm>;
defm INT_PTX_ATOM_OR_GEN_32_USE_G : F_ATOMIC_2<Int32Regs, ".global", ".b32",
".or", atomic_load_or_32_gen, i32imm, imm>;
defm INT_PTX_ATOM_OR_S_32 : F_ATOMIC_2<Int32Regs, ".shared", ".b32", ".or",
atomic_load_or_32_s, i32imm, imm>;
defm INT_PTX_ATOM_OR_G_64 : F_ATOMIC_2<Int64Regs, ".global", ".b64", ".or",
atomic_load_or_64_g, i64imm, imm>;
defm INT_PTX_ATOM_OR_GEN_64 : F_ATOMIC_2<Int64Regs, "", ".b64", ".or",
atomic_load_or_64_gen, i64imm, imm>;
defm INT_PTX_ATOM_OR_GEN_64_USE_G : F_ATOMIC_2<Int64Regs, ".global", ".b64",
".or", atomic_load_or_64_gen, i64imm, imm>;
defm INT_PTX_ATOM_OR_S_64 : F_ATOMIC_2<Int64Regs, ".shared", ".b64", ".or",
atomic_load_or_64_s, i64imm, imm>;
// atom_xor
def atomic_load_xor_32_g: ATOMIC_GLOBAL_CHK<(ops node:$a, node:$b),
(atomic_load_xor_32 node:$a, node:$b)>;
def atomic_load_xor_32_s: ATOMIC_SHARED_CHK<(ops node:$a, node:$b),
(atomic_load_xor_32 node:$a, node:$b)>;
def atomic_load_xor_32_gen: ATOMIC_GENERIC_CHK<(ops node:$a, node:$b),
(atomic_load_xor_32 node:$a, node:$b)>;
def atomic_load_xor_64_g: ATOMIC_GLOBAL_CHK<(ops node:$a, node:$b),
(atomic_load_xor_64 node:$a, node:$b)>;
def atomic_load_xor_64_s: ATOMIC_SHARED_CHK<(ops node:$a, node:$b),
(atomic_load_xor_64 node:$a, node:$b)>;
def atomic_load_xor_64_gen: ATOMIC_GENERIC_CHK<(ops node:$a, node:$b),
(atomic_load_xor_64 node:$a, node:$b)>;
defm INT_PTX_ATOM_XOR_G_32 : F_ATOMIC_2<Int32Regs, ".global", ".b32", ".xor",
atomic_load_xor_32_g, i32imm, imm>;
defm INT_PTX_ATOM_XOR_S_32 : F_ATOMIC_2<Int32Regs, ".shared", ".b32", ".xor",
atomic_load_xor_32_s, i32imm, imm>;
defm INT_PTX_ATOM_XOR_GEN_32 : F_ATOMIC_2<Int32Regs, "", ".b32", ".xor",
atomic_load_xor_32_gen, i32imm, imm>;
defm INT_PTX_ATOM_XOR_GEN_32_USE_G : F_ATOMIC_2<Int32Regs, ".global", ".b32",
".xor", atomic_load_xor_32_gen, i32imm, imm>;
defm INT_PTX_ATOM_XOR_G_64 : F_ATOMIC_2<Int64Regs, ".global", ".b64", ".xor",
atomic_load_xor_64_g, i64imm, imm>;
defm INT_PTX_ATOM_XOR_S_64 : F_ATOMIC_2<Int64Regs, ".shared", ".b64", ".xor",
atomic_load_xor_64_s, i64imm, imm>;
defm INT_PTX_ATOM_XOR_GEN_64 : F_ATOMIC_2<Int64Regs, "", ".b64", ".xor",
atomic_load_xor_64_gen, i64imm, imm>;
defm INT_PTX_ATOM_XOR_GEN_64_USE_G : F_ATOMIC_2<Int64Regs, ".global", ".b64",
".xor", atomic_load_xor_64_gen, i64imm, imm>;
// atom_cas
def atomic_cmp_swap_32_g: ATOMIC_GLOBAL_CHK<(ops node:$a, node:$b, node:$c),
(atomic_cmp_swap_32 node:$a, node:$b, node:$c)>;
def atomic_cmp_swap_32_s: ATOMIC_SHARED_CHK<(ops node:$a, node:$b, node:$c),
(atomic_cmp_swap_32 node:$a, node:$b, node:$c)>;
def atomic_cmp_swap_32_gen: ATOMIC_GENERIC_CHK<(ops node:$a, node:$b, node:$c),
(atomic_cmp_swap_32 node:$a, node:$b, node:$c)>;
def atomic_cmp_swap_64_g: ATOMIC_GLOBAL_CHK<(ops node:$a, node:$b, node:$c),
(atomic_cmp_swap_64 node:$a, node:$b, node:$c)>;
def atomic_cmp_swap_64_s: ATOMIC_SHARED_CHK<(ops node:$a, node:$b, node:$c),
(atomic_cmp_swap_64 node:$a, node:$b, node:$c)>;
def atomic_cmp_swap_64_gen: ATOMIC_GENERIC_CHK<(ops node:$a, node:$b, node:$c),
(atomic_cmp_swap_64 node:$a, node:$b, node:$c)>;
defm INT_PTX_ATOM_CAS_G_32 : F_ATOMIC_3<Int32Regs, ".global", ".b32", ".cas",
atomic_cmp_swap_32_g, i32imm>;
defm INT_PTX_ATOM_CAS_S_32 : F_ATOMIC_3<Int32Regs, ".shared", ".b32", ".cas",
atomic_cmp_swap_32_s, i32imm>;
defm INT_PTX_ATOM_CAS_GEN_32 : F_ATOMIC_3<Int32Regs, "", ".b32", ".cas",
atomic_cmp_swap_32_gen, i32imm>;
defm INT_PTX_ATOM_CAS_GEN_32_USE_G : F_ATOMIC_3<Int32Regs, ".global", ".b32",
".cas", atomic_cmp_swap_32_gen, i32imm>;
defm INT_PTX_ATOM_CAS_G_64 : F_ATOMIC_3<Int64Regs, ".global", ".b64", ".cas",
atomic_cmp_swap_64_g, i64imm>;
defm INT_PTX_ATOM_CAS_S_64 : F_ATOMIC_3<Int64Regs, ".shared", ".b64", ".cas",
atomic_cmp_swap_64_s, i64imm>;
defm INT_PTX_ATOM_CAS_GEN_64 : F_ATOMIC_3<Int64Regs, "", ".b64", ".cas",
atomic_cmp_swap_64_gen, i64imm>;
defm INT_PTX_ATOM_CAS_GEN_64_USE_G : F_ATOMIC_3<Int64Regs, ".global", ".b64",
".cas", atomic_cmp_swap_64_gen, i64imm>;
// Support for scoped atomic operations. Matches
// int_nvvm_atomic_{op}_{space}_{type}_{scope}
// and converts it into the appropriate instruction.
// NOTE: not all possible combinations are implemented
// 'space' is limited to generic as it's the only one needed to support CUDA.
// 'scope' = 'gpu' is default and is handled by regular atomic instructions.
class ATOM23_impl<string AsmStr, NVPTXRegClass regclass, list<Predicate> Preds,
dag ins, dag Operands>
: NVPTXInst<(outs regclass:$result), ins,
AsmStr,
[(set regclass:$result, Operands)]>,
Requires<Preds>;
// Define instruction variants for all addressing modes.
multiclass ATOM2P_impl<string AsmStr, Intrinsic Intr,
NVPTXRegClass regclass, Operand ImmType,
SDNode Imm, ValueType ImmTy,
list<Predicate> Preds> {
let AddedComplexity = 1 in {
def : ATOM23_impl<AsmStr, regclass, Preds,
(ins Int32Regs:$src, regclass:$b),
(Intr Int32Regs:$src, regclass:$b)>;
def : ATOM23_impl<AsmStr, regclass, Preds,
(ins Int64Regs:$src, regclass:$b),
(Intr Int64Regs:$src, regclass:$b)>;
}
// tablegen can't infer argument types from Intrinsic (though it can
// from Instruction) so we have to enforce specific type on
// immediates via explicit cast to ImmTy.
def : ATOM23_impl<AsmStr, regclass, Preds,
(ins Int32Regs:$src, ImmType:$b),
(Intr Int32Regs:$src, (ImmTy Imm:$b))>;
def : ATOM23_impl<AsmStr, regclass, Preds,
(ins Int64Regs:$src, ImmType:$b),
(Intr Int64Regs:$src, (ImmTy Imm:$b))>;
}
multiclass ATOM3P_impl<string AsmStr, Intrinsic Intr,
NVPTXRegClass regclass, Operand ImmType,
SDNode Imm, ValueType ImmTy,
list<Predicate> Preds> {
// Variants for register/immediate permutations of $b and $c
let AddedComplexity = 2 in {
def : ATOM23_impl<AsmStr, regclass, Preds,
(ins Int32Regs:$src, regclass:$b, regclass:$c),
(Intr Int32Regs:$src, regclass:$b, regclass:$c)>;
def : ATOM23_impl<AsmStr, regclass, Preds,
(ins Int64Regs:$src, regclass:$b, regclass:$c),
(Intr Int64Regs:$src, regclass:$b, regclass:$c)>;
}
let AddedComplexity = 1 in {
def : ATOM23_impl<AsmStr, regclass, Preds,
(ins Int32Regs:$src, ImmType:$b, regclass:$c),
(Intr Int32Regs:$src, (ImmTy Imm:$b), regclass:$c)>;
def : ATOM23_impl<AsmStr, regclass, Preds,
(ins Int64Regs:$src, ImmType:$b, regclass:$c),
(Intr Int64Regs:$src, (ImmTy Imm:$b), regclass:$c)>;
def : ATOM23_impl<AsmStr, regclass, Preds,
(ins Int32Regs:$src, regclass:$b, ImmType:$c),
(Intr Int32Regs:$src, regclass:$b, (ImmTy Imm:$c))>;
def : ATOM23_impl<AsmStr, regclass, Preds,
(ins Int64Regs:$src, regclass:$b, ImmType:$c),
(Intr Int64Regs:$src, regclass:$b, (ImmTy Imm:$c))>;
}
def : ATOM23_impl<AsmStr, regclass, Preds,
(ins Int32Regs:$src, ImmType:$b, ImmType:$c),
(Intr Int32Regs:$src, (ImmTy Imm:$b), (ImmTy Imm:$c))>;
def : ATOM23_impl<AsmStr, regclass, Preds,
(ins Int64Regs:$src, ImmType:$b, ImmType:$c),
(Intr Int64Regs:$src, (ImmTy Imm:$b), (ImmTy Imm:$c))>;
}
// Constructs instrinsic name and instruction asm strings.
multiclass ATOM2N_impl<string OpStr, string IntTypeStr, string TypeStr,
string ScopeStr, string SpaceStr,
NVPTXRegClass regclass, Operand ImmType, SDNode Imm,
ValueType ImmTy, list<Predicate> Preds> {
defm : ATOM2P_impl<"atom" # !if(!eq(SpaceStr, "gen"), "", "." # SpaceStr)
# !if(!eq(ScopeStr, "gpu"), "", "." # ScopeStr)
# "." # OpStr # "." # TypeStr
# " \t$result, [$src], $b;",
!cast<Intrinsic>(
"int_nvvm_atomic_" # OpStr
# "_" # SpaceStr # "_" # IntTypeStr
# !if(!eq(ScopeStr,""), "", "_" # ScopeStr)),
regclass, ImmType, Imm, ImmTy, Preds>;
}
multiclass ATOM3N_impl<string OpStr, string IntTypeStr, string TypeStr,
string ScopeStr, string SpaceStr,
NVPTXRegClass regclass, Operand ImmType, SDNode Imm,
ValueType ImmTy, list<Predicate> Preds> {
defm : ATOM3P_impl<"atom" # !if(!eq(SpaceStr, "gen"), "", "." # SpaceStr)
# !if(!eq(ScopeStr, "gpu"), "", "." # ScopeStr)
# "." # OpStr # "." # TypeStr
# " \t$result, [$src], $b, $c;",
!cast<Intrinsic>(
"int_nvvm_atomic_" # OpStr
# "_" # SpaceStr # "_" # IntTypeStr
# !if(!eq(ScopeStr,""), "", "_" # ScopeStr)),
regclass, ImmType, Imm, ImmTy, Preds>;
}
// Constructs variants for different address spaces.
// For now we only need variants for generic space pointers.
multiclass ATOM2A_impl<string OpStr, string IntTypeStr, string TypeStr,
string ScopeStr, NVPTXRegClass regclass, Operand ImmType,
SDNode Imm, ValueType ImmTy, list<Predicate> Preds> {
defm _gen_ : ATOM2N_impl<OpStr, IntTypeStr, TypeStr, ScopeStr, "gen",
regclass, ImmType, Imm, ImmTy, Preds>;
}
multiclass ATOM3A_impl<string OpStr, string IntTypeStr, string TypeStr,
string ScopeStr, NVPTXRegClass regclass, Operand ImmType,
SDNode Imm, ValueType ImmTy, list<Predicate> Preds> {
defm _gen_ : ATOM3N_impl<OpStr, IntTypeStr, TypeStr, ScopeStr, "gen",
regclass, ImmType, Imm, ImmTy, Preds>;
}
// Constructs variants for different scopes of atomic op.
multiclass ATOM2S_impl<string OpStr, string IntTypeStr, string TypeStr,
NVPTXRegClass regclass, Operand ImmType, SDNode Imm,
ValueType ImmTy, list<Predicate> Preds> {
// .gpu scope is default and is currently covered by existing
// atomics w/o explicitly specified scope.
defm _cta : ATOM2A_impl<OpStr, IntTypeStr, TypeStr, "cta",
regclass, ImmType, Imm, ImmTy,
!listconcat(Preds,[hasAtomScope])>;
defm _sys : ATOM2A_impl<OpStr, IntTypeStr, TypeStr, "sys",
regclass, ImmType, Imm, ImmTy,
!listconcat(Preds,[hasAtomScope])>;
}
multiclass ATOM3S_impl<string OpStr, string IntTypeStr, string TypeStr,
NVPTXRegClass regclass, Operand ImmType, SDNode Imm, ValueType ImmTy,
list<Predicate> Preds> {
// No need to define ".gpu"-scoped atomics. They do the same thing
// as the regular, non-scoped atomics defined elsewhere.
defm _cta : ATOM3A_impl<OpStr, IntTypeStr, TypeStr, "cta",
regclass, ImmType, Imm, ImmTy,
!listconcat(Preds,[hasAtomScope])>;
defm _sys : ATOM3A_impl<OpStr, IntTypeStr, TypeStr, "sys",
regclass, ImmType, Imm, ImmTy,
!listconcat(Preds,[hasAtomScope])>;
}
// atom.add
multiclass ATOM2_add_impl<string OpStr> {
defm _s32 : ATOM2S_impl<OpStr, "i", "s32", Int32Regs, i32imm, imm, i32, []>;
defm _u32 : ATOM2S_impl<OpStr, "i", "u32", Int32Regs, i32imm, imm, i32, []>;
defm _u64 : ATOM2S_impl<OpStr, "i", "u64", Int64Regs, i64imm, imm, i64, []>;
defm _f32 : ATOM2S_impl<OpStr, "f", "f32", Float32Regs, f32imm, fpimm, f32,
[]>;
defm _f64 : ATOM2S_impl<OpStr, "f", "f64", Float64Regs, f64imm, fpimm, f64,
[hasAtomAddF64]>;
}
// atom.{and,or,xor}
multiclass ATOM2_bitwise_impl<string OpStr> {
defm _b32 : ATOM2S_impl<OpStr, "i", "b32", Int32Regs, i32imm, imm, i32, []>;
defm _b64 : ATOM2S_impl<OpStr, "i", "b64", Int64Regs, i64imm, imm, i64,
[hasAtomBitwise64]>;
}
// atom.exch
multiclass ATOM2_exch_impl<string OpStr> {
defm _b32 : ATOM2S_impl<OpStr, "i", "b32", Int32Regs, i32imm, imm, i32, []>;
defm _b64 : ATOM2S_impl<OpStr, "i", "b64", Int64Regs, i64imm, imm, i64, []>;
}
// atom.{min,max}
multiclass ATOM2_minmax_impl<string OpStr> {
defm _s32 : ATOM2S_impl<OpStr, "i", "s32", Int32Regs, i32imm, imm, i32, []>;
defm _u32 : ATOM2S_impl<OpStr, "i", "u32", Int32Regs, i32imm, imm, i32, []>;
defm _s64 : ATOM2S_impl<OpStr, "i", "s64", Int64Regs, i64imm, imm, i64,
[hasAtomMinMax64]>;
defm _u64 : ATOM2S_impl<OpStr, "i", "u64", Int64Regs, i64imm, imm, i64,
[hasAtomMinMax64]>;
}
// atom.{inc,dec}
multiclass ATOM2_incdec_impl<string OpStr> {
defm _u32 : ATOM2S_impl<OpStr, "i", "u32", Int32Regs, i32imm, imm, i32, []>;
}
// atom.cas
multiclass ATOM3_cas_impl<string OpStr> {
defm _b32 : ATOM3S_impl<OpStr, "i", "b32", Int32Regs, i32imm, imm, i32, []>;
defm _b64 : ATOM3S_impl<OpStr, "i", "b64", Int64Regs, i64imm, imm, i64, []>;
}
defm INT_PTX_SATOM_ADD : ATOM2_add_impl<"add">;
defm INT_PTX_SATOM_AND : ATOM2_bitwise_impl<"and">;
defm INT_PTX_SATOM_CAS : ATOM3_cas_impl<"cas">;
defm INT_PTX_SATOM_DEC : ATOM2_incdec_impl<"dec">;
defm INT_PTX_SATOM_EXCH: ATOM2_exch_impl<"exch">;
defm INT_PTX_SATOM_INC : ATOM2_incdec_impl<"inc">;
defm INT_PTX_SATOM_MAX : ATOM2_minmax_impl<"max">;
defm INT_PTX_SATOM_MIN : ATOM2_minmax_impl<"min">;
defm INT_PTX_SATOM_OR : ATOM2_bitwise_impl<"or">;
defm INT_PTX_SATOM_XOR : ATOM2_bitwise_impl<"xor">;
//-----------------------------------
// Support for ldu on sm_20 or later
//-----------------------------------
// Don't annotate ldu instructions as mayLoad, as they load from memory that is
// read-only in a kernel.
// Scalar
multiclass LDU_G<string TyStr, NVPTXRegClass regclass> {
def areg: NVPTXInst<(outs regclass:$result), (ins Int32Regs:$src),
!strconcat("ldu.global.", TyStr),
[]>, Requires<[hasLDU]>;
def areg64: NVPTXInst<(outs regclass:$result), (ins Int64Regs:$src),
!strconcat("ldu.global.", TyStr),
[]>, Requires<[hasLDU]>;
def avar: NVPTXInst<(outs regclass:$result), (ins imemAny:$src),
!strconcat("ldu.global.", TyStr),
[]>, Requires<[hasLDU]>;
def ari : NVPTXInst<(outs regclass:$result), (ins MEMri:$src),
!strconcat("ldu.global.", TyStr),
[]>, Requires<[hasLDU]>;
def ari64 : NVPTXInst<(outs regclass:$result), (ins MEMri64:$src),
!strconcat("ldu.global.", TyStr),
[]>, Requires<[hasLDU]>;
}
defm INT_PTX_LDU_GLOBAL_i8 : LDU_G<"u8 \t$result, [$src];", Int16Regs>;
defm INT_PTX_LDU_GLOBAL_i16 : LDU_G<"u16 \t$result, [$src];", Int16Regs>;
defm INT_PTX_LDU_GLOBAL_i32 : LDU_G<"u32 \t$result, [$src];", Int32Regs>;
defm INT_PTX_LDU_GLOBAL_i64 : LDU_G<"u64 \t$result, [$src];", Int64Regs>;
defm INT_PTX_LDU_GLOBAL_f16 : LDU_G<"b16 \t$result, [$src];", Float16Regs>;
defm INT_PTX_LDU_GLOBAL_f16x2 : LDU_G<"b32 \t$result, [$src];", Float16x2Regs>;
defm INT_PTX_LDU_GLOBAL_f32 : LDU_G<"f32 \t$result, [$src];", Float32Regs>;
defm INT_PTX_LDU_GLOBAL_f64 : LDU_G<"f64 \t$result, [$src];", Float64Regs>;
defm INT_PTX_LDU_GLOBAL_p32 : LDU_G<"u32 \t$result, [$src];", Int32Regs>;
defm INT_PTX_LDU_GLOBAL_p64 : LDU_G<"u64 \t$result, [$src];", Int64Regs>;
// vector
// Elementized vector ldu
multiclass VLDU_G_ELE_V2<string TyStr, NVPTXRegClass regclass> {
def _areg32: NVPTXInst<(outs regclass:$dst1, regclass:$dst2),
(ins Int32Regs:$src),
!strconcat("ldu.global.", TyStr), []>;
def _areg64: NVPTXInst<(outs regclass:$dst1, regclass:$dst2),
(ins Int64Regs:$src),
!strconcat("ldu.global.", TyStr), []>;
def _ari32: NVPTXInst<(outs regclass:$dst1, regclass:$dst2),
(ins MEMri:$src),
!strconcat("ldu.global.", TyStr), []>;
def _ari64: NVPTXInst<(outs regclass:$dst1, regclass:$dst2),
(ins MEMri64:$src),
!strconcat("ldu.global.", TyStr), []>;
def _avar: NVPTXInst<(outs regclass:$dst1, regclass:$dst2),
(ins imemAny:$src),
!strconcat("ldu.global.", TyStr), []>;
}
multiclass VLDU_G_ELE_V4<string TyStr, NVPTXRegClass regclass> {
def _areg32: NVPTXInst<(outs regclass:$dst1, regclass:$dst2, regclass:$dst3,
regclass:$dst4), (ins Int32Regs:$src),
!strconcat("ldu.global.", TyStr), []>;
def _areg64: NVPTXInst<(outs regclass:$dst1, regclass:$dst2, regclass:$dst3,
regclass:$dst4), (ins Int64Regs:$src),
!strconcat("ldu.global.", TyStr), []>;
def _ari32: NVPTXInst<(outs regclass:$dst1, regclass:$dst2, regclass:$dst3,
regclass:$dst4), (ins MEMri:$src),
!strconcat("ldu.global.", TyStr), []>;
def _ari64: NVPTXInst<(outs regclass:$dst1, regclass:$dst2, regclass:$dst3,
regclass:$dst4), (ins MEMri64:$src),
!strconcat("ldu.global.", TyStr), []>;
def _avar: NVPTXInst<(outs regclass:$dst1, regclass:$dst2, regclass:$dst3,
regclass:$dst4), (ins imemAny:$src),
!strconcat("ldu.global.", TyStr), []>;
}
defm INT_PTX_LDU_G_v2i8_ELE
: VLDU_G_ELE_V2<"v2.u8 \t{{$dst1, $dst2}}, [$src];", Int16Regs>;
defm INT_PTX_LDU_G_v2i16_ELE
: VLDU_G_ELE_V2<"v2.u16 \t{{$dst1, $dst2}}, [$src];", Int16Regs>;
defm INT_PTX_LDU_G_v2i32_ELE
: VLDU_G_ELE_V2<"v2.u32 \t{{$dst1, $dst2}}, [$src];", Int32Regs>;
defm INT_PTX_LDU_G_v2f16_ELE
: VLDU_G_ELE_V2<"v2.b16 \t{{$dst1, $dst2}}, [$src];", Float16Regs>;
defm INT_PTX_LDU_G_v2f16x2_ELE
: VLDU_G_ELE_V2<"v2.b32 \t{{$dst1, $dst2}}, [$src];", Float16x2Regs>;
defm INT_PTX_LDU_G_v2f32_ELE
: VLDU_G_ELE_V2<"v2.f32 \t{{$dst1, $dst2}}, [$src];", Float32Regs>;
defm INT_PTX_LDU_G_v2i64_ELE
: VLDU_G_ELE_V2<"v2.u64 \t{{$dst1, $dst2}}, [$src];", Int64Regs>;
defm INT_PTX_LDU_G_v2f64_ELE
: VLDU_G_ELE_V2<"v2.f64 \t{{$dst1, $dst2}}, [$src];", Float64Regs>;
defm INT_PTX_LDU_G_v4i8_ELE
: VLDU_G_ELE_V4<"v4.u8 \t{{$dst1, $dst2, $dst3, $dst4}}, [$src];", Int16Regs>;
defm INT_PTX_LDU_G_v4i16_ELE
: VLDU_G_ELE_V4<"v4.u16 \t{{$dst1, $dst2, $dst3, $dst4}}, [$src];",
Int16Regs>;
defm INT_PTX_LDU_G_v4i32_ELE
: VLDU_G_ELE_V4<"v4.u32 \t{{$dst1, $dst2, $dst3, $dst4}}, [$src];",
Int32Regs>;
defm INT_PTX_LDU_G_v4f16_ELE
: VLDU_G_ELE_V4<"v4.b16 \t{{$dst1, $dst2, $dst3, $dst4}}, [$src];",
Float16Regs>;
defm INT_PTX_LDU_G_v4f16x2_ELE
: VLDU_G_ELE_V4<"v4.b32 \t{{$dst1, $dst2, $dst3, $dst4}}, [$src];",
Float16x2Regs>;
defm INT_PTX_LDU_G_v4f32_ELE
: VLDU_G_ELE_V4<"v4.f32 \t{{$dst1, $dst2, $dst3, $dst4}}, [$src];",
Float32Regs>;
//-----------------------------------
// Support for ldg on sm_35 or later
//-----------------------------------
// Don't annotate ld.global.nc as mayLoad, because these loads go through the
// non-coherent texture cache, and therefore the values read must be read-only
// during the lifetime of the kernel.
multiclass LDG_G<string TyStr, NVPTXRegClass regclass> {
def areg: NVPTXInst<(outs regclass:$result), (ins Int32Regs:$src),
!strconcat("ld.global.nc.", TyStr),
[]>, Requires<[hasLDG]>;
def areg64: NVPTXInst<(outs regclass:$result), (ins Int64Regs:$src),
!strconcat("ld.global.nc.", TyStr),
[]>, Requires<[hasLDG]>;
def avar: NVPTXInst<(outs regclass:$result), (ins imemAny:$src),
!strconcat("ld.global.nc.", TyStr),
[]>, Requires<[hasLDG]>;
def ari : NVPTXInst<(outs regclass:$result), (ins MEMri:$src),
!strconcat("ld.global.nc.", TyStr),
[]>, Requires<[hasLDG]>;
def ari64 : NVPTXInst<(outs regclass:$result), (ins MEMri64:$src),
!strconcat("ld.global.nc.", TyStr),
[]>, Requires<[hasLDG]>;
}
defm INT_PTX_LDG_GLOBAL_i8
: LDG_G<"u8 \t$result, [$src];", Int16Regs>;
defm INT_PTX_LDG_GLOBAL_i16
: LDG_G<"u16 \t$result, [$src];", Int16Regs>;
defm INT_PTX_LDG_GLOBAL_i32
: LDG_G<"u32 \t$result, [$src];", Int32Regs>;
defm INT_PTX_LDG_GLOBAL_i64
: LDG_G<"u64 \t$result, [$src];", Int64Regs>;
defm INT_PTX_LDG_GLOBAL_f16
: LDG_G<"b16 \t$result, [$src];", Float16Regs>;
defm INT_PTX_LDG_GLOBAL_f16x2
: LDG_G<"b32 \t$result, [$src];", Float16x2Regs>;
defm INT_PTX_LDG_GLOBAL_f32
: LDG_G<"f32 \t$result, [$src];", Float32Regs>;
defm INT_PTX_LDG_GLOBAL_f64
: LDG_G<"f64 \t$result, [$src];", Float64Regs>;
defm INT_PTX_LDG_GLOBAL_p32
: LDG_G<"u32 \t$result, [$src];", Int32Regs>;
defm INT_PTX_LDG_GLOBAL_p64
: LDG_G<"u64 \t$result, [$src];", Int64Regs>;
// vector
// Elementized vector ldg
multiclass VLDG_G_ELE_V2<string TyStr, NVPTXRegClass regclass> {
def _areg32: NVPTXInst<(outs regclass:$dst1, regclass:$dst2),
(ins Int32Regs:$src),
!strconcat("ld.global.nc.", TyStr), []>;
def _areg64: NVPTXInst<(outs regclass:$dst1, regclass:$dst2),
(ins Int64Regs:$src),
!strconcat("ld.global.nc.", TyStr), []>;
def _ari32: NVPTXInst<(outs regclass:$dst1, regclass:$dst2),
(ins MEMri:$src),
!strconcat("ld.global.nc.", TyStr), []>;
def _ari64: NVPTXInst<(outs regclass:$dst1, regclass:$dst2),
(ins MEMri64:$src),
!strconcat("ld.global.nc.", TyStr), []>;
def _avar: NVPTXInst<(outs regclass:$dst1, regclass:$dst2),
(ins imemAny:$src),
!strconcat("ld.global.nc.", TyStr), []>;
}
multiclass VLDG_G_ELE_V4<string TyStr, NVPTXRegClass regclass> {
def _areg32: NVPTXInst<(outs regclass:$dst1, regclass:$dst2, regclass:$dst3,
regclass:$dst4), (ins Int32Regs:$src),
!strconcat("ld.global.nc.", TyStr), []>;
def _areg64: NVPTXInst<(outs regclass:$dst1, regclass:$dst2, regclass:$dst3,
regclass:$dst4), (ins Int64Regs:$src),
!strconcat("ld.global.nc.", TyStr), []>;
def _ari32: NVPTXInst<(outs regclass:$dst1, regclass:$dst2, regclass:$dst3,
regclass:$dst4), (ins MEMri:$src),
!strconcat("ld.global.nc.", TyStr), []>;
def _ari64: NVPTXInst<(outs regclass:$dst1, regclass:$dst2, regclass:$dst3,
regclass:$dst4), (ins MEMri64:$src),
!strconcat("ld.global.nc.", TyStr), []>;
def _avar: NVPTXInst<(outs regclass:$dst1, regclass:$dst2, regclass:$dst3,
regclass:$dst4), (ins imemAny:$src),
!strconcat("ld.global.nc.", TyStr), []>;
}
// FIXME: 8-bit LDG should be fixed once LDG/LDU nodes are made into proper loads.
defm INT_PTX_LDG_G_v2i8_ELE
: VLDG_G_ELE_V2<"v2.u8 \t{{$dst1, $dst2}}, [$src];", Int16Regs>;
defm INT_PTX_LDG_G_v2i16_ELE
: VLDG_G_ELE_V2<"v2.u16 \t{{$dst1, $dst2}}, [$src];", Int16Regs>;
defm INT_PTX_LDG_G_v2i32_ELE
: VLDG_G_ELE_V2<"v2.u32 \t{{$dst1, $dst2}}, [$src];", Int32Regs>;
defm INT_PTX_LDG_G_v2f16_ELE
: VLDG_G_ELE_V2<"v2.b16 \t{{$dst1, $dst2}}, [$src];", Float16Regs>;
defm INT_PTX_LDG_G_v2f16x2_ELE
: VLDG_G_ELE_V2<"v2.b32 \t{{$dst1, $dst2}}, [$src];", Float16x2Regs>;
defm INT_PTX_LDG_G_v2f32_ELE
: VLDG_G_ELE_V2<"v2.f32 \t{{$dst1, $dst2}}, [$src];", Float32Regs>;
defm INT_PTX_LDG_G_v2i64_ELE
: VLDG_G_ELE_V2<"v2.u64 \t{{$dst1, $dst2}}, [$src];", Int64Regs>;
defm INT_PTX_LDG_G_v2f64_ELE
: VLDG_G_ELE_V2<"v2.f64 \t{{$dst1, $dst2}}, [$src];", Float64Regs>;
defm INT_PTX_LDG_G_v4i8_ELE
: VLDG_G_ELE_V4<"v4.u8 \t{{$dst1, $dst2, $dst3, $dst4}}, [$src];", Int16Regs>;
defm INT_PTX_LDG_G_v4i16_ELE
: VLDG_G_ELE_V4<"v4.u16 \t{{$dst1, $dst2, $dst3, $dst4}}, [$src];", Int16Regs>;
defm INT_PTX_LDG_G_v4i32_ELE
: VLDG_G_ELE_V4<"v4.u32 \t{{$dst1, $dst2, $dst3, $dst4}}, [$src];", Int32Regs>;
defm INT_PTX_LDG_G_v4f16_ELE
: VLDG_G_ELE_V4<"v4.b16 \t{{$dst1, $dst2, $dst3, $dst4}}, [$src];", Float16Regs>;
defm INT_PTX_LDG_G_v4f16x2_ELE
: VLDG_G_ELE_V4<"v4.b32 \t{{$dst1, $dst2, $dst3, $dst4}}, [$src];", Float16x2Regs>;
defm INT_PTX_LDG_G_v4f32_ELE
: VLDG_G_ELE_V4<"v4.f32 \t{{$dst1, $dst2, $dst3, $dst4}}, [$src];", Float32Regs>;
multiclass NG_TO_G<string Str, Intrinsic Intrin> {
def _yes : NVPTXInst<(outs Int32Regs:$result), (ins Int32Regs:$src),
!strconcat("cvta.", Str, ".u32 \t$result, $src;"),
[(set Int32Regs:$result, (Intrin Int32Regs:$src))]>;
def _yes_64 : NVPTXInst<(outs Int64Regs:$result), (ins Int64Regs:$src),
!strconcat("cvta.", Str, ".u64 \t$result, $src;"),
[(set Int64Regs:$result, (Intrin Int64Regs:$src))]>;
def _yes_6432 : NVPTXInst<(outs Int64Regs:$result), (ins Int32Regs:$src),
"{{ .reg .b64 %tmp;\n\t"
#" cvt.u64.u32 \t%tmp, $src;\n\t"
#" cvta." # Str # ".u64 \t$result, %tmp; }}",
[(set Int64Regs:$result, (Intrin Int32Regs:$src))]>,
Requires<[useShortPtr]>;
}
multiclass G_TO_NG<string Str, Intrinsic Intrin> {
def _yes : NVPTXInst<(outs Int32Regs:$result), (ins Int32Regs:$src),
!strconcat("cvta.to.", Str, ".u32 \t$result, $src;"),
[(set Int32Regs:$result, (Intrin Int32Regs:$src))]>;
def _yes_64 : NVPTXInst<(outs Int64Regs:$result), (ins Int64Regs:$src),
!strconcat("cvta.to.", Str, ".u64 \t$result, $src;"),
[(set Int64Regs:$result, (Intrin Int64Regs:$src))]>;
def _yes_3264 : NVPTXInst<(outs Int32Regs:$result), (ins Int64Regs:$src),
"{{ .reg .b64 %tmp;\n\t"
#" cvta.to." # Str # ".u64 \t%tmp, $src;\n\t"
#" cvt.u32.u64 \t$result, %tmp; }}",
[(set Int32Regs:$result, (Intrin Int64Regs:$src))]>,
Requires<[useShortPtr]>;
}
defm cvta_local : NG_TO_G<"local", int_nvvm_ptr_local_to_gen>;
defm cvta_shared : NG_TO_G<"shared", int_nvvm_ptr_shared_to_gen>;
defm cvta_global : NG_TO_G<"global", int_nvvm_ptr_global_to_gen>;
defm cvta_const : NG_TO_G<"const", int_nvvm_ptr_constant_to_gen>;
defm cvta_to_local : G_TO_NG<"local", int_nvvm_ptr_gen_to_local>;
defm cvta_to_shared : G_TO_NG<"shared", int_nvvm_ptr_gen_to_shared>;
defm cvta_to_global : G_TO_NG<"global", int_nvvm_ptr_gen_to_global>;
defm cvta_to_const : G_TO_NG<"const", int_nvvm_ptr_gen_to_constant>;
// nvvm.ptr.gen.to.param
def nvvm_ptr_gen_to_param : NVPTXInst<(outs Int32Regs:$result),
(ins Int32Regs:$src),
"mov.u32 \t$result, $src;",
[(set Int32Regs:$result,
(int_nvvm_ptr_gen_to_param Int32Regs:$src))]>;
def nvvm_ptr_gen_to_param_64 : NVPTXInst<(outs Int64Regs:$result),
(ins Int64Regs:$src),
"mov.u64 \t$result, $src;",
[(set Int64Regs:$result,
(int_nvvm_ptr_gen_to_param Int64Regs:$src))]>;
// nvvm.move intrinsicc
def nvvm_move_i16 : NVPTXInst<(outs Int16Regs:$r), (ins Int16Regs:$s),
"mov.b16 \t$r, $s;",
[(set Int16Regs:$r,
(int_nvvm_move_i16 Int16Regs:$s))]>;
def nvvm_move_i32 : NVPTXInst<(outs Int32Regs:$r), (ins Int32Regs:$s),
"mov.b32 \t$r, $s;",
[(set Int32Regs:$r,
(int_nvvm_move_i32 Int32Regs:$s))]>;
def nvvm_move_i64 : NVPTXInst<(outs Int64Regs:$r), (ins Int64Regs:$s),
"mov.b64 \t$r, $s;",
[(set Int64Regs:$r,
(int_nvvm_move_i64 Int64Regs:$s))]>;
def nvvm_move_float : NVPTXInst<(outs Float32Regs:$r), (ins Float32Regs:$s),
"mov.f32 \t$r, $s;",
[(set Float32Regs:$r,
(int_nvvm_move_float Float32Regs:$s))]>;
def nvvm_move_double : NVPTXInst<(outs Float64Regs:$r), (ins Float64Regs:$s),
"mov.f64 \t$r, $s;",
[(set Float64Regs:$r,
(int_nvvm_move_double Float64Regs:$s))]>;
def nvvm_move_ptr32 : NVPTXInst<(outs Int32Regs:$r), (ins Int32Regs:$s),
"mov.u32 \t$r, $s;",
[(set Int32Regs:$r,
(int_nvvm_move_ptr Int32Regs:$s))]>;
def nvvm_move_ptr64 : NVPTXInst<(outs Int64Regs:$r), (ins Int64Regs:$s),
"mov.u64 \t$r, $s;",
[(set Int64Regs:$r,
(int_nvvm_move_ptr Int64Regs:$s))]>;
// @TODO: Are these actually needed, or will we always just see symbols
// copied to registers first?
/*def nvvm_move_sym32 : NVPTXInst<(outs Int32Regs:$r), (ins imem:$s),
"mov.u32 \t$r, $s;",
[(set Int32Regs:$r,
(int_nvvm_move_ptr texternalsym:$s))]>;
def nvvm_move_sym64 : NVPTXInst<(outs Int64Regs:$r), (ins imem:$s),
"mov.u64 \t$r, $s;",
[(set Int64Regs:$r,
(int_nvvm_move_ptr texternalsym:$s))]>;*/
// MoveParam %r1, param
// ptr_local_to_gen %r2, %r1
// ptr_gen_to_local %r3, %r2
// ->
// mov %r1, param
// @TODO: Revisit this. There is a type
// contradiction between iPTRAny and iPTR for the addr defs, so the move_sym
// instructions are not currently defined. However, we can use the ptr
// variants and the asm printer will do the right thing.
def : Pat<(i64 (int_nvvm_ptr_gen_to_local (int_nvvm_ptr_local_to_gen
(MoveParam texternalsym:$src)))),
(nvvm_move_ptr64 texternalsym:$src)>;
def : Pat<(i32 (int_nvvm_ptr_gen_to_local (int_nvvm_ptr_local_to_gen
(MoveParam texternalsym:$src)))),
(nvvm_move_ptr32 texternalsym:$src)>;
def texsurf_handles
: NVPTXInst<(outs Int64Regs:$result), (ins imem:$src),
"mov.u64 \t$result, $src;", []>;
//-----------------------------------
// Compiler Error Warn
// - Just ignore them in codegen
//-----------------------------------
def INT_NVVM_COMPILER_WARN_32 : NVPTXInst<(outs), (ins Int32Regs:$a),
"// llvm.nvvm.compiler.warn()",
[(int_nvvm_compiler_warn Int32Regs:$a)]>;
def INT_NVVM_COMPILER_WARN_64 : NVPTXInst<(outs), (ins Int64Regs:$a),
"// llvm.nvvm.compiler.warn()",
[(int_nvvm_compiler_warn Int64Regs:$a)]>;
def INT_NVVM_COMPILER_ERROR_32 : NVPTXInst<(outs), (ins Int32Regs:$a),
"// llvm.nvvm.compiler.error()",
[(int_nvvm_compiler_error Int32Regs:$a)]>;
def INT_NVVM_COMPILER_ERROR_64 : NVPTXInst<(outs), (ins Int64Regs:$a),
"// llvm.nvvm.compiler.error()",
[(int_nvvm_compiler_error Int64Regs:$a)]>;
// isspacep
def ISSPACEP_CONST_32
: NVPTXInst<(outs Int1Regs:$d), (ins Int32Regs:$a),
"isspacep.const \t$d, $a;",
[(set Int1Regs:$d, (int_nvvm_isspacep_const Int32Regs:$a))]>,
Requires<[hasPTX31]>;
def ISSPACEP_CONST_64
: NVPTXInst<(outs Int1Regs:$d), (ins Int64Regs:$a),
"isspacep.const \t$d, $a;",
[(set Int1Regs:$d, (int_nvvm_isspacep_const Int64Regs:$a))]>,
Requires<[hasPTX31]>;
def ISSPACEP_GLOBAL_32
: NVPTXInst<(outs Int1Regs:$d), (ins Int32Regs:$a),
"isspacep.global \t$d, $a;",
[(set Int1Regs:$d, (int_nvvm_isspacep_global Int32Regs:$a))]>;
def ISSPACEP_GLOBAL_64
: NVPTXInst<(outs Int1Regs:$d), (ins Int64Regs:$a),
"isspacep.global \t$d, $a;",
[(set Int1Regs:$d, (int_nvvm_isspacep_global Int64Regs:$a))]>;
def ISSPACEP_LOCAL_32
: NVPTXInst<(outs Int1Regs:$d), (ins Int32Regs:$a),
"isspacep.local \t$d, $a;",
[(set Int1Regs:$d, (int_nvvm_isspacep_local Int32Regs:$a))]>;
def ISSPACEP_LOCAL_64
: NVPTXInst<(outs Int1Regs:$d), (ins Int64Regs:$a),
"isspacep.local \t$d, $a;",
[(set Int1Regs:$d, (int_nvvm_isspacep_local Int64Regs:$a))]>;
def ISSPACEP_SHARED_32
: NVPTXInst<(outs Int1Regs:$d), (ins Int32Regs:$a),
"isspacep.shared \t$d, $a;",
[(set Int1Regs:$d, (int_nvvm_isspacep_shared Int32Regs:$a))]>;
def ISSPACEP_SHARED_64
: NVPTXInst<(outs Int1Regs:$d), (ins Int64Regs:$a),
"isspacep.shared \t$d, $a;",
[(set Int1Regs:$d, (int_nvvm_isspacep_shared Int64Regs:$a))]>;
// Special register reads
def MOV_SPECIAL : NVPTXInst<(outs Int32Regs:$d),
(ins SpecialRegs:$r),
"mov.b32 \t$d, $r;", []>;
def : Pat<(int_nvvm_read_ptx_sreg_envreg0), (MOV_SPECIAL ENVREG0)>;
def : Pat<(int_nvvm_read_ptx_sreg_envreg1), (MOV_SPECIAL ENVREG1)>;
def : Pat<(int_nvvm_read_ptx_sreg_envreg2), (MOV_SPECIAL ENVREG2)>;
def : Pat<(int_nvvm_read_ptx_sreg_envreg3), (MOV_SPECIAL ENVREG3)>;
def : Pat<(int_nvvm_read_ptx_sreg_envreg4), (MOV_SPECIAL ENVREG4)>;
def : Pat<(int_nvvm_read_ptx_sreg_envreg5), (MOV_SPECIAL ENVREG5)>;
def : Pat<(int_nvvm_read_ptx_sreg_envreg6), (MOV_SPECIAL ENVREG6)>;
def : Pat<(int_nvvm_read_ptx_sreg_envreg7), (MOV_SPECIAL ENVREG7)>;
def : Pat<(int_nvvm_read_ptx_sreg_envreg8), (MOV_SPECIAL ENVREG8)>;
def : Pat<(int_nvvm_read_ptx_sreg_envreg9), (MOV_SPECIAL ENVREG9)>;
def : Pat<(int_nvvm_read_ptx_sreg_envreg10), (MOV_SPECIAL ENVREG10)>;
def : Pat<(int_nvvm_read_ptx_sreg_envreg11), (MOV_SPECIAL ENVREG11)>;
def : Pat<(int_nvvm_read_ptx_sreg_envreg12), (MOV_SPECIAL ENVREG12)>;
def : Pat<(int_nvvm_read_ptx_sreg_envreg13), (MOV_SPECIAL ENVREG13)>;
def : Pat<(int_nvvm_read_ptx_sreg_envreg14), (MOV_SPECIAL ENVREG14)>;
def : Pat<(int_nvvm_read_ptx_sreg_envreg15), (MOV_SPECIAL ENVREG15)>;
def : Pat<(int_nvvm_read_ptx_sreg_envreg16), (MOV_SPECIAL ENVREG16)>;
def : Pat<(int_nvvm_read_ptx_sreg_envreg17), (MOV_SPECIAL ENVREG17)>;
def : Pat<(int_nvvm_read_ptx_sreg_envreg18), (MOV_SPECIAL ENVREG18)>;
def : Pat<(int_nvvm_read_ptx_sreg_envreg19),