multiclass avx512_subvec_broadcast_rm<bits<8> opc, string OpcodeStr,
X86VectorVTInfo _Dst, X86VectorVTInfo _Src> {
- let mayLoad = 1 in {
- def rm : AVX5128I<opc, MRMSrcMem, (outs _Dst.RC:$dst), (ins _Src.MemOp:$src),
- !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
- [(set _Dst.RC:$dst,
- (_Dst.VT (X86SubVBroadcast
- (_Src.VT (bitconvert (_Src.LdFrag addr:$src))))))]>, EVEX;
- def rmk : AVX5128I<opc, MRMSrcMem, (outs _Dst.RC:$dst), (ins _Dst.KRCWM:$mask,
- _Src.MemOp:$src),
- !strconcat(OpcodeStr,
- "\t{$src, ${dst} {${mask}}|${dst} {${mask}}, $src}"),
- []>, EVEX, EVEX_K;
- def rmkz : AVX5128I<opc, MRMSrcMem, (outs _Dst.RC:$dst), (ins _Dst.KRCWM:$mask,
- _Src.MemOp:$src),
- !strconcat(OpcodeStr,
- "\t{$src, ${dst} {${mask}} {z}|${dst} {${mask}} {z}, $src}"),
- []>, EVEX, EVEX_KZ;
- }
+ let mayLoad = 1 in
+ defm rm : AVX512_maskable<opc, MRMSrcMem, _Dst, (outs _Dst.RC:$dst),
+ (ins _Src.MemOp:$src), OpcodeStr, "$src", "$src",
+ (_Dst.VT (X86SubVBroadcast
+ (_Src.VT (bitconvert (_Src.LdFrag addr:$src)))))>,
+ AVX5128IBase, EVEX;
}
defm VBROADCASTI32X4 : avx512_subvec_broadcast_rm<0x5a, "vbroadcasti32x4",
(EXTRACT_SUBREG
(AND32ri (KMOVWrk (COPY_TO_REGCLASS VK1:$src, VK16)), (i32 1)),
sub_16bit)>;
- def : Pat<(v16i1 (scalar_to_vector VK1:$src)),
- (COPY_TO_REGCLASS VK1:$src, VK16)>;
- def : Pat<(v8i1 (scalar_to_vector VK1:$src)),
- (COPY_TO_REGCLASS VK1:$src, VK8)>;
-}
-let Predicates = [HasBWI] in {
- def : Pat<(v32i1 (scalar_to_vector VK1:$src)),
- (COPY_TO_REGCLASS VK1:$src, VK32)>;
- def : Pat<(v64i1 (scalar_to_vector VK1:$src)),
- (COPY_TO_REGCLASS VK1:$src, VK64)>;
}
+def : Pat<(v16i1 (scalar_to_vector VK1:$src)),
+ (COPY_TO_REGCLASS VK1:$src, VK16)>;
+def : Pat<(v8i1 (scalar_to_vector VK1:$src)),
+ (COPY_TO_REGCLASS VK1:$src, VK8)>;
+def : Pat<(v4i1 (scalar_to_vector VK1:$src)),
+ (COPY_TO_REGCLASS VK1:$src, VK4)>;
+def : Pat<(v2i1 (scalar_to_vector VK1:$src)),
+ (COPY_TO_REGCLASS VK1:$src, VK2)>;
+def : Pat<(v32i1 (scalar_to_vector VK1:$src)),
+ (COPY_TO_REGCLASS VK1:$src, VK32)>;
+def : Pat<(v64i1 (scalar_to_vector VK1:$src)),
+ (COPY_TO_REGCLASS VK1:$src, VK64)>;
// With AVX-512 only, 8-bit mask is promoted to 16-bit mask.
def : Pat<(v16i1 (extract_subvector (v32i1 VK32:$src), (iPTR 0))),
(v16i1 (COPY_TO_REGCLASS VK32:$src, VK16))>;
+def : Pat<(v16i1 (extract_subvector (v32i1 VK32:$src), (iPTR 16))),
+ (v16i1 (COPY_TO_REGCLASS (KSHIFTRDri VK32:$src, (i8 16)), VK16))>;
+
def : Pat<(v32i1 (extract_subvector (v64i1 VK64:$src), (iPTR 0))),
(v32i1 (COPY_TO_REGCLASS VK64:$src, VK32))>;
def : Pat<(v4i1 (extract_subvector (v8i1 VK8:$src), (iPTR 0))),
(v4i1 (COPY_TO_REGCLASS VK8:$src, VK4))>;
+
def : Pat<(v2i1 (extract_subvector (v8i1 VK8:$src), (iPTR 0))),
(v2i1 (COPY_TO_REGCLASS VK8:$src, VK2))>;
avx512vl_i64_info>, VEX_W;
}
+// Use 512bit version to implement 128/256 bit in case NoVLX.
+multiclass avx512_var_shift_w_lowering<AVX512VLVectorVTInfo _, SDNode OpNode> {
+ let Predicates = [HasBWI, NoVLX] in {
+ def : Pat<(_.info256.VT (OpNode (_.info256.VT _.info256.RC:$src1),
+ (_.info256.VT _.info256.RC:$src2))),
+ (EXTRACT_SUBREG
+ (!cast<Instruction>(NAME#"WZrr")
+ (INSERT_SUBREG (_.info512.VT (IMPLICIT_DEF)), VR256X:$src1, sub_ymm),
+ (INSERT_SUBREG (_.info512.VT (IMPLICIT_DEF)), VR256X:$src2, sub_ymm)),
+ sub_ymm)>;
+
+ def : Pat<(_.info128.VT (OpNode (_.info128.VT _.info128.RC:$src1),
+ (_.info128.VT _.info128.RC:$src2))),
+ (EXTRACT_SUBREG
+ (!cast<Instruction>(NAME#"WZrr")
+ (INSERT_SUBREG (_.info512.VT (IMPLICIT_DEF)), VR128X:$src1, sub_xmm),
+ (INSERT_SUBREG (_.info512.VT (IMPLICIT_DEF)), VR128X:$src2, sub_xmm)),
+ sub_xmm)>;
+ }
+}
+
multiclass avx512_var_shift_w<bits<8> opc, string OpcodeStr,
SDNode OpNode> {
let Predicates = [HasBWI] in
}
defm VPSLLV : avx512_var_shift_types<0x47, "vpsllv", shl>,
- avx512_var_shift_w<0x12, "vpsllvw", shl>;
+ avx512_var_shift_w<0x12, "vpsllvw", shl>,
+ avx512_var_shift_w_lowering<avx512vl_i16_info, shl>;
defm VPSRAV : avx512_var_shift_types<0x46, "vpsrav", sra>,
- avx512_var_shift_w<0x11, "vpsravw", sra>;
+ avx512_var_shift_w<0x11, "vpsravw", sra>,
+ avx512_var_shift_w_lowering<avx512vl_i16_info, sra>;
defm VPSRLV : avx512_var_shift_types<0x45, "vpsrlv", srl>,
- avx512_var_shift_w<0x10, "vpsrlvw", srl>;
+ avx512_var_shift_w<0x10, "vpsrlvw", srl>,
+ avx512_var_shift_w_lowering<avx512vl_i16_info, srl>;
defm VPRORV : avx512_var_shift_types<0x14, "vprorv", rotr>;
defm VPROLV : avx512_var_shift_types<0x15, "vprolv", rotl>;
defm NAME: avx512_shift_rmi_sizes<OpcImm, MRMSrcReg, MRMSrcMem, OpcodeStr,
X86VPermilpi, _>,
EVEX, AVX512AIi8Base, EVEX_CD8<_.info128.EltSize, CD8VF>;
-
- let isCodeGenOnly = 1 in {
- // lowering implementation with the alternative types
- defm NAME#_I: avx512_permil_vec_common<OpcodeStr, OpcVar, Ctrl, Ctrl>;
- defm NAME#_I: avx512_shift_rmi_sizes<OpcImm, MRMSrcReg, MRMSrcMem,
- OpcodeStr, X86VPermilpi, Ctrl>,
- EVEX, AVX512AIi8Base, EVEX_CD8<_.info128.EltSize, CD8VF>;
- }
}
defm VPERMILPS : avx512_permil<"vpermilps", 0x04, 0x0C, avx512vl_f32_info,
defm VRSQRT14 : avx512_fp14_p_vl_all<0x4E, "vrsqrt14", X86frsqrt>;
defm VRCP14 : avx512_fp14_p_vl_all<0x4C, "vrcp14", X86frcp>;
-def : Pat <(v16f32 (int_x86_avx512_rsqrt14_ps_512 (v16f32 VR512:$src),
- (bc_v16f32 (v16i32 immAllZerosV)), (i16 -1))),
- (VRSQRT14PSZr VR512:$src)>;
-def : Pat <(v8f64 (int_x86_avx512_rsqrt14_pd_512 (v8f64 VR512:$src),
- (bc_v8f64 (v16i32 immAllZerosV)), (i8 -1))),
- (VRSQRT14PDZr VR512:$src)>;
-
-def : Pat <(v16f32 (int_x86_avx512_rcp14_ps_512 (v16f32 VR512:$src),
- (bc_v16f32 (v16i32 immAllZerosV)), (i16 -1))),
- (VRCP14PSZr VR512:$src)>;
-def : Pat <(v8f64 (int_x86_avx512_rcp14_pd_512 (v8f64 VR512:$src),
- (bc_v8f64 (v16i32 immAllZerosV)), (i8 -1))),
- (VRCP14PDZr VR512:$src)>;
-
/// avx512_fp28_s rcp28ss, rcp28sd, rsqrt28ss, rsqrt28sd
multiclass avx512_fp28_s<bits<8> opc, string OpcodeStr,X86VectorVTInfo _,
SDNode OpNode> {
multiclass convert_vector_to_mask_common<bits<8> opc, X86VectorVTInfo _, string OpcodeStr > {
def rr : AVX512XS8I<opc, MRMSrcReg, (outs _.KRC:$dst), (ins _.RC:$src),
!strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
- [(set _.KRC:$dst, (trunc (_.VT _.RC:$src)))]>, EVEX;
+ [(set _.KRC:$dst, (X86cvt2mask (_.VT _.RC:$src)))]>, EVEX;
}
multiclass avx512_convert_vector_to_mask<bits<8> opc, string OpcodeStr,
defm VSHUFI64X2 : avx512_shuff_packed_128<"vshufi64x2",avx512vl_i64_info, 0x43>,
AVX512AIi8Base, EVEX_4V, EVEX_CD8<64, CD8VF>, VEX_W;
-multiclass avx512_valign<string OpcodeStr, AVX512VLVectorVTInfo VTInfo_I,
- AVX512VLVectorVTInfo VTInfo_FP>{
+multiclass avx512_valign<string OpcodeStr, AVX512VLVectorVTInfo VTInfo_I> {
defm NAME: avx512_common_3Op_imm8<OpcodeStr, VTInfo_I, 0x03, X86VAlign>,
AVX512AIi8Base, EVEX_4V;
- let isCodeGenOnly = 1 in {
- defm NAME#_FP: avx512_common_3Op_imm8<OpcodeStr, VTInfo_FP, 0x03, X86VAlign>,
- AVX512AIi8Base, EVEX_4V;
- }
}
-defm VALIGND: avx512_valign<"valignd", avx512vl_i32_info, avx512vl_f32_info>,
+defm VALIGND: avx512_valign<"valignd", avx512vl_i32_info>,
EVEX_CD8<32, CD8VF>;
-defm VALIGNQ: avx512_valign<"valignq", avx512vl_i64_info, avx512vl_f64_info>,
+defm VALIGNQ: avx512_valign<"valignq", avx512vl_i64_info>,
EVEX_CD8<64, CD8VF>, VEX_W;
multiclass avx512_vpalign_lowering<X86VectorVTInfo _ , list<Predicate> p>{
multiclass avx512_replicate<bits<8> opc, string OpcodeStr, SDNode OpNode>{
defm NAME: avx512_unary_rm_vl<opc, OpcodeStr, OpNode, avx512vl_f32_info,
HasAVX512>, XS;
- let isCodeGenOnly = 1 in
- defm NAME#_I: avx512_unary_rm_vl<opc, OpcodeStr, OpNode, avx512vl_i32_info,
- HasAVX512>, XS;
}
defm VMOVSHDUP : avx512_replicate<0x16, "vmovshdup", X86Movshdup>;
multiclass avx512_movddup<bits<8> opc, string OpcodeStr, SDNode OpNode>{
defm NAME: avx512_movddup_common<opc, OpcodeStr, OpNode,
avx512vl_f64_info>, XD, VEX_W;
- let isCodeGenOnly = 1 in
- defm NAME#_I: avx512_movddup_common<opc, OpcodeStr, OpNode,
- avx512vl_i64_info>;
}
defm VMOVDDUP : avx512_movddup<0x12, "vmovddup", X86Movddup>;
defm NAME: avx512_common_3Op_imm8<OpcodeStr, VTInfo_FP, 0xC6, X86Shufp>,
EVEX_CD8<VTInfo_FP.info512.EltSize, CD8VF>,
AVX512AIi8Base, EVEX_4V;
- let isCodeGenOnly = 1 in {
- defm NAME#_I: avx512_common_3Op_imm8<OpcodeStr, VTInfo_I, 0xC6, X86Shufp>,
- EVEX_CD8<VTInfo_I.info512.EltSize, CD8VF>,
- AVX512AIi8Base, EVEX_4V;
- }
}
defm VSHUFPS: avx512_shufp<"vshufps", avx512vl_i32_info, avx512vl_f32_info>, PS;
projects / oota-llvm.git / blobdiff