//===----------------------------------------------------------------------===//
// -- VPERM2I - 3 source operands form --
-multiclass avx512_perm_3src<bits<8> opc, string OpcodeStr,
+multiclass avx512_perm_i<bits<8> opc, string OpcodeStr,
SDNode OpNode, X86VectorVTInfo _> {
let Constraints = "$src1 = $dst" in {
defm rr: AVX512_maskable_3src<opc, MRMSrcReg, _, (outs _.RC:$dst),
EVEX_4V, AVX5128IBase;
}
}
-multiclass avx512_perm_3src_mb<bits<8> opc, string OpcodeStr,
+multiclass avx512_perm_i_mb<bits<8> opc, string OpcodeStr,
SDNode OpNode, X86VectorVTInfo _> {
let mayLoad = 1, Constraints = "$src1 = $dst" in
defm rmb: AVX512_maskable_3src<opc, MRMSrcMem, _, (outs _.RC:$dst),
AVX5128IBase, EVEX_4V, EVEX_B;
}
-multiclass avx512_perm_3src_sizes<bits<8> opc, string OpcodeStr,
+multiclass avx512_perm_i_sizes<bits<8> opc, string OpcodeStr,
SDNode OpNode, AVX512VLVectorVTInfo VTInfo> {
- let Predicates = [HasAVX512] in
- defm NAME: avx512_perm_3src<opc, OpcodeStr, OpNode, VTInfo.info512>,
- avx512_perm_3src_mb<opc, OpcodeStr, OpNode, VTInfo.info512>, EVEX_V512;
+ defm NAME: avx512_perm_i<opc, OpcodeStr, OpNode, VTInfo.info512>,
+ avx512_perm_i_mb<opc, OpcodeStr, OpNode, VTInfo.info512>, EVEX_V512;
let Predicates = [HasVLX] in {
- defm NAME#128: avx512_perm_3src<opc, OpcodeStr, OpNode, VTInfo.info128>,
- avx512_perm_3src_mb<opc, OpcodeStr, OpNode, VTInfo.info128>,
- EVEX_V128;
- defm NAME#256: avx512_perm_3src<opc, OpcodeStr, OpNode, VTInfo.info256>,
- avx512_perm_3src_mb<opc, OpcodeStr, OpNode, VTInfo.info256>,
- EVEX_V256;
+ defm NAME#128: avx512_perm_i<opc, OpcodeStr, OpNode, VTInfo.info128>,
+ avx512_perm_i_mb<opc, OpcodeStr, OpNode, VTInfo.info128>, EVEX_V128;
+ defm NAME#256: avx512_perm_i<opc, OpcodeStr, OpNode, VTInfo.info256>,
+ avx512_perm_i_mb<opc, OpcodeStr, OpNode, VTInfo.info256>, EVEX_V256;
}
}
-multiclass avx512_perm_3src_sizes_w<bits<8> opc, string OpcodeStr,
- SDNode OpNode, AVX512VLVectorVTInfo VTInfo> {
+
+multiclass avx512_perm_i_sizes_w<bits<8> opc, string OpcodeStr,
+ SDNode OpNode, AVX512VLVectorVTInfo VTInfo> {
let Predicates = [HasBWI] in
- defm NAME: avx512_perm_3src<opc, OpcodeStr, OpNode, VTInfo.info512>,
- avx512_perm_3src_mb<opc, OpcodeStr, OpNode, VTInfo.info512>,
- EVEX_V512;
+ defm NAME: avx512_perm_i<opc, OpcodeStr, OpNode, VTInfo.info512>, EVEX_V512;
let Predicates = [HasBWI, HasVLX] in {
- defm NAME#128: avx512_perm_3src<opc, OpcodeStr, OpNode, VTInfo.info128>,
- avx512_perm_3src_mb<opc, OpcodeStr, OpNode, VTInfo.info128>,
- EVEX_V128;
- defm NAME#256: avx512_perm_3src<opc, OpcodeStr, OpNode, VTInfo.info256>,
- avx512_perm_3src_mb<opc, OpcodeStr, OpNode, VTInfo.info256>,
- EVEX_V256;
- }
-}
-defm VPERMI2D : avx512_perm_3src_sizes<0x76, "vpermi2d", X86VPermiv3,
- avx512vl_i32_info>, EVEX_CD8<32, CD8VF>;
-defm VPERMI2Q : avx512_perm_3src_sizes<0x76, "vpermi2q", X86VPermiv3,
- avx512vl_i64_info>, VEX_W, EVEX_CD8<64, CD8VF>;
-defm VPERMI2PS : avx512_perm_3src_sizes<0x77, "vpermi2ps", X86VPermiv3,
- avx512vl_f32_info>, EVEX_CD8<32, CD8VF>;
-defm VPERMI2PD : avx512_perm_3src_sizes<0x77, "vpermi2pd", X86VPermiv3,
- avx512vl_f64_info>, VEX_W, EVEX_CD8<64, CD8VF>;
-
-defm VPERMT2D : avx512_perm_3src_sizes<0x7E, "vpermt2d", X86VPermv3,
- avx512vl_i32_info>, EVEX_CD8<32, CD8VF>;
-defm VPERMT2Q : avx512_perm_3src_sizes<0x7E, "vpermt2q", X86VPermv3,
- avx512vl_i64_info>, VEX_W, EVEX_CD8<64, CD8VF>;
-defm VPERMT2PS : avx512_perm_3src_sizes<0x7F, "vpermt2ps", X86VPermv3,
- avx512vl_f32_info>, EVEX_CD8<32, CD8VF>;
-defm VPERMT2PD : avx512_perm_3src_sizes<0x7F, "vpermt2pd", X86VPermv3,
- avx512vl_f64_info>, VEX_W, EVEX_CD8<64, CD8VF>;
-
-defm VPERMT2W : avx512_perm_3src_sizes_w<0x7D, "vpermt2w", X86VPermv3,
- avx512vl_i16_info>, VEX_W, EVEX_CD8<16, CD8VF>;
-defm VPERMI2W : avx512_perm_3src_sizes_w<0x75, "vpermi2w", X86VPermiv3,
- avx512vl_i16_info>, VEX_W, EVEX_CD8<16, CD8VF>;
+ defm NAME#128: avx512_perm_i<opc, OpcodeStr, OpNode, VTInfo.info128>, EVEX_V128;
+ defm NAME#256: avx512_perm_i<opc, OpcodeStr, OpNode, VTInfo.info256>, EVEX_V256;
+ }
+}
+
+defm VPERMI2D : avx512_perm_i_sizes<0x76, "vpermi2d", X86VPermi2X,
+ avx512vl_i32_info>, EVEX_CD8<32, CD8VF>;
+defm VPERMI2Q : avx512_perm_i_sizes<0x76, "vpermi2q", X86VPermi2X,
+ avx512vl_i64_info>, VEX_W, EVEX_CD8<64, CD8VF>;
+defm VPERMI2W : avx512_perm_i_sizes_w<0x75, "vpermi2w", X86VPermi2X,
+ avx512vl_i16_info>, VEX_W, EVEX_CD8<16, CD8VF>;
+defm VPERMI2PS : avx512_perm_i_sizes<0x77, "vpermi2ps", X86VPermi2X,
+ avx512vl_f32_info>, EVEX_CD8<32, CD8VF>;
+defm VPERMI2PD : avx512_perm_i_sizes<0x77, "vpermi2pd", X86VPermi2X,
+ avx512vl_f64_info>, VEX_W, EVEX_CD8<64, CD8VF>;
+
+// VPERMT
+multiclass avx512_perm_t<bits<8> opc, string OpcodeStr,
+ X86VectorVTInfo _, X86VectorVTInfo IdxVT> {
+let Constraints = "$src1 = $dst" in {
+ defm rr: AVX512_maskable_3src<opc, MRMSrcReg, _, (outs _.RC:$dst),
+ (ins IdxVT.RC:$src2, _.RC:$src3),
+ OpcodeStr, "$src3, $src2", "$src2, $src3",
+ (_.VT (X86VPermt2 _.RC:$src1, IdxVT.RC:$src2, _.RC:$src3))>, EVEX_4V,
+ AVX5128IBase;
+
+ let mayLoad = 1 in
+ defm rm: AVX512_maskable_3src<opc, MRMSrcMem, _, (outs _.RC:$dst),
+ (ins IdxVT.RC:$src2, _.MemOp:$src3),
+ OpcodeStr, "$src3, $src2", "$src2, $src3",
+ (_.VT (X86VPermt2 _.RC:$src1, IdxVT.RC:$src2,
+ (bitconvert (_.LdFrag addr:$src3))))>,
+ EVEX_4V, AVX5128IBase;
+ }
+}
+multiclass avx512_perm_t_mb<bits<8> opc, string OpcodeStr,
+ X86VectorVTInfo _, X86VectorVTInfo IdxVT> {
+ let mayLoad = 1, Constraints = "$src1 = $dst" in
+ defm rmb: AVX512_maskable_3src<opc, MRMSrcMem, _, (outs _.RC:$dst),
+ (ins IdxVT.RC:$src2, _.ScalarMemOp:$src3),
+ OpcodeStr, !strconcat("${src3}", _.BroadcastStr,", $src2"),
+ !strconcat("$src2, ${src3}", _.BroadcastStr ),
+ (_.VT (X86VPermt2 _.RC:$src1,
+ IdxVT.RC:$src2,(_.VT (X86VBroadcast (_.ScalarLdFrag addr:$src3)))))>,
+ AVX5128IBase, EVEX_4V, EVEX_B;
+}
+
+multiclass avx512_perm_t_sizes<bits<8> opc, string OpcodeStr,
+ AVX512VLVectorVTInfo VTInfo,
+ AVX512VLVectorVTInfo ShuffleMask> {
+ defm NAME: avx512_perm_t<opc, OpcodeStr, VTInfo.info512,
+ ShuffleMask.info512>,
+ avx512_perm_t_mb<opc, OpcodeStr, VTInfo.info512,
+ ShuffleMask.info512>, EVEX_V512;
+ let Predicates = [HasVLX] in {
+ defm NAME#128: avx512_perm_t<opc, OpcodeStr, VTInfo.info128,
+ ShuffleMask.info128>,
+ avx512_perm_t_mb<opc, OpcodeStr, VTInfo.info128,
+ ShuffleMask.info128>, EVEX_V128;
+ defm NAME#256: avx512_perm_t<opc, OpcodeStr, VTInfo.info256,
+ ShuffleMask.info256>,
+ avx512_perm_t_mb<opc, OpcodeStr, VTInfo.info256,
+ ShuffleMask.info256>, EVEX_V256;
+ }
+}
+
+multiclass avx512_perm_t_sizes_w<bits<8> opc, string OpcodeStr,
+ AVX512VLVectorVTInfo VTInfo,
+ AVX512VLVectorVTInfo Idx> {
+ let Predicates = [HasBWI] in
+ defm NAME: avx512_perm_t<opc, OpcodeStr, VTInfo.info512,
+ Idx.info512>, EVEX_V512;
+ let Predicates = [HasBWI, HasVLX] in {
+ defm NAME#128: avx512_perm_t<opc, OpcodeStr, VTInfo.info128,
+ Idx.info128>, EVEX_V128;
+ defm NAME#256: avx512_perm_t<opc, OpcodeStr, VTInfo.info256,
+ Idx.info256>, EVEX_V256;
+ }
+}
+
+defm VPERMT2D : avx512_perm_t_sizes<0x7E, "vpermt2d",
+ avx512vl_i32_info, avx512vl_i32_info>, EVEX_CD8<32, CD8VF>;
+defm VPERMT2Q : avx512_perm_t_sizes<0x7E, "vpermt2q",
+ avx512vl_i64_info, avx512vl_i64_info>, VEX_W, EVEX_CD8<64, CD8VF>;
+defm VPERMT2W : avx512_perm_t_sizes_w<0x7D, "vpermt2w",
+ avx512vl_i16_info, avx512vl_i16_info>, VEX_W, EVEX_CD8<16, CD8VF>;
+defm VPERMT2PS : avx512_perm_t_sizes<0x7F, "vpermt2ps",
+ avx512vl_f32_info, avx512vl_i32_info>, EVEX_CD8<32, CD8VF>;
+defm VPERMT2PD : avx512_perm_t_sizes<0x7F, "vpermt2pd",
+ avx512vl_f64_info, avx512vl_i64_info>, VEX_W, EVEX_CD8<64, CD8VF>;
//===----------------------------------------------------------------------===//
// AVX-512 - BLEND using mask
defm VCVTUQQ2PS : avx512_cvtqq2ps<0x7A, "vcvtuqq2ps", uint_to_fp,
X86VUlongToFpRnd>, VEX_W, XD, EVEX_CD8<64, CD8VF>;
-let Predicates = [NoVLX] in {
+let Predicates = [HasAVX512, NoVLX] in {
def : Pat<(v8i32 (fp_to_uint (v8f32 VR256X:$src1))),
(EXTRACT_SUBREG (v16i32 (VCVTTPS2UDQZrr
(v16f32 (SUBREG_TO_REG (i32 0), VR256X:$src1, sub_ymm)))), sub_ymm)>;
multiclass avx512_psadbw_packed<bits<8> opc, SDNode OpNode,
- string OpcodeStr, X86VectorVTInfo _src>{
+ string OpcodeStr, X86VectorVTInfo _dst,
+ X86VectorVTInfo _src>{
def rr : AVX512BI<opc, MRMSrcReg,
- (outs _src.RC:$dst), (ins _src.RC:$src1, _src.RC:$src2),
+ (outs _dst.RC:$dst), (ins _src.RC:$src1, _src.RC:$src2),
!strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
- [(set _src.RC:$dst,(_src.VT
- (OpNode _src.RC:$src1, _src.RC:$src2)))]>;
+ [(set _dst.RC:$dst,(_dst.VT
+ (OpNode (_src.VT _src.RC:$src1),
+ (_src.VT _src.RC:$src2))))]>;
let mayLoad = 1 in
def rm : AVX512BI<opc, MRMSrcMem,
- (outs _src.RC:$dst), (ins _src.RC:$src1, _src.MemOp:$src2),
+ (outs _dst.RC:$dst), (ins _src.RC:$src1, _src.MemOp:$src2),
!strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
- [(set _src.RC:$dst,(_src.VT
- (OpNode _src.RC:$src1,
- (_src.VT (bitconvert
+ [(set _dst.RC:$dst,(_dst.VT
+ (OpNode (_src.VT _src.RC:$src1),
+ (_src.VT (bitconvert
(_src.LdFrag addr:$src2))))))]>;
}
multiclass avx512_psadbw_packed_all<bits<8> opc, SDNode OpNode,
string OpcodeStr, Predicate prd> {
let Predicates = [prd] in
- defm Z512 : avx512_psadbw_packed<opc, OpNode, OpcodeStr, v64i8_info>,
- EVEX_V512;
+ defm Z512 : avx512_psadbw_packed<opc, OpNode, OpcodeStr, v8i64_info,
+ v64i8_info>, EVEX_V512;
let Predicates = [prd, HasVLX] in {
- defm Z256 : avx512_psadbw_packed<opc, OpNode, OpcodeStr, v32i8x_info>,
- EVEX_V256;
- defm Z128 : avx512_psadbw_packed<opc, OpNode, OpcodeStr, v16i8x_info>,
- EVEX_V128;
+ defm Z256 : avx512_psadbw_packed<opc, OpNode, OpcodeStr, v4i64x_info,
+ v32i8x_info>, EVEX_V256;
+ defm Z128 : avx512_psadbw_packed<opc, OpNode, OpcodeStr, v2i64x_info,
+ v16i8x_info>, EVEX_V128;
}
}
projects / oota-llvm.git / blobdiff