X-Git-Url: http://plrg.eecs.uci.edu/git/?p=oota-llvm.git;a=blobdiff_plain;f=lib%2FTarget%2FX86%2FX86InstrAVX512.td;h=f7e5d9c7b526df18cb119f36fb7b525dd6441641;hp=6176af6af0b45a4cff7b4c8e364b2803e2d1fadc;hb=d34ae248c07aac9e00b6d505c4d56e95de66c244;hpb=e20dfebf8715a249c330b289740d4882d1396ae9 diff --git a/lib/Target/X86/X86InstrAVX512.td b/lib/Target/X86/X86InstrAVX512.td index 6176af6af0b..f7e5d9c7b52 100644 --- a/lib/Target/X86/X86InstrAVX512.td +++ b/lib/Target/X86/X86InstrAVX512.td @@ -276,6 +276,22 @@ multiclass AVX512_maskable_3src O, Format F, X86VectorVTInfo _, OpcodeStr, AttSrcAsm, IntelSrcAsm, RHS, (vselect _.KRCWM:$mask, RHS, _.RC:$src1)>; +// Similar to AVX512_maskable_3rc but in this case the input VT for the tied +// operand differs from the output VT. This requires a bitconvert on +// the preserved vector going into the vselect. +multiclass AVX512_maskable_3src_cast O, Format F, X86VectorVTInfo OutVT, + X86VectorVTInfo InVT, + dag Outs, dag NonTiedIns, string OpcodeStr, + string AttSrcAsm, string IntelSrcAsm, + dag RHS> : + AVX512_maskable_common; + multiclass AVX512_maskable_3src_scalar O, Format F, X86VectorVTInfo _, dag Outs, dag NonTiedIns, string OpcodeStr, string AttSrcAsm, string IntelSrcAsm, @@ -1136,88 +1152,155 @@ defm VPBROADCASTMB2Q : avx512_mask_broadcast<0x2A, "vpbroadcastmb2q", avx512vl_i64_info, VK8>, VEX_W; //===----------------------------------------------------------------------===// -// -- VPERM2I - 3 source operands form -- -multiclass avx512_perm_3src opc, string OpcodeStr, - SDNode OpNode, X86VectorVTInfo _> { +// -- VPERMI2 - 3 source operands form -- +multiclass avx512_perm_i opc, string OpcodeStr, + X86VectorVTInfo _, X86VectorVTInfo IdxVT> { let Constraints = "$src1 = $dst" in { - defm rr: AVX512_maskable_3src, EVEX_4V, + (_.VT (X86VPermi2X IdxVT.RC:$src1, _.RC:$src2, _.RC:$src3))>, EVEX_4V, AVX5128IBase; let mayLoad = 1 in - defm rm: AVX512_maskable_3src, EVEX_4V, AVX5128IBase; } } -multiclass avx512_perm_3src_mb opc, string OpcodeStr, - SDNode OpNode, X86VectorVTInfo _> { +multiclass avx512_perm_i_mb opc, string OpcodeStr, + X86VectorVTInfo _, X86VectorVTInfo IdxVT> { let mayLoad = 1, Constraints = "$src1 = $dst" in - defm rmb: AVX512_maskable_3src, AVX5128IBase, EVEX_4V, EVEX_B; } -multiclass avx512_perm_3src_sizes opc, string OpcodeStr, - SDNode OpNode, AVX512VLVectorVTInfo VTInfo> { - let Predicates = [HasAVX512] in - defm NAME: avx512_perm_3src, - avx512_perm_3src_mb, EVEX_V512; +multiclass avx512_perm_i_sizes opc, string OpcodeStr, + AVX512VLVectorVTInfo VTInfo, + AVX512VLVectorVTInfo ShuffleMask> { + defm NAME: avx512_perm_i, + avx512_perm_i_mb, EVEX_V512; + let Predicates = [HasVLX] in { + defm NAME#128: avx512_perm_i, + avx512_perm_i_mb, EVEX_V128; + defm NAME#256: avx512_perm_i, + avx512_perm_i_mb, EVEX_V256; + } +} + +multiclass avx512_perm_i_sizes_w opc, string OpcodeStr, + AVX512VLVectorVTInfo VTInfo, + AVX512VLVectorVTInfo Idx> { + let Predicates = [HasBWI] in + defm NAME: avx512_perm_i, EVEX_V512; + let Predicates = [HasBWI, HasVLX] in { + defm NAME#128: avx512_perm_i, EVEX_V128; + defm NAME#256: avx512_perm_i, EVEX_V256; + } +} + +defm VPERMI2D : avx512_perm_i_sizes<0x76, "vpermi2d", + avx512vl_i32_info, avx512vl_i32_info>, EVEX_CD8<32, CD8VF>; +defm VPERMI2Q : avx512_perm_i_sizes<0x76, "vpermi2q", + avx512vl_i64_info, avx512vl_i64_info>, VEX_W, EVEX_CD8<64, CD8VF>; +defm VPERMI2W : avx512_perm_i_sizes_w<0x75, "vpermi2w", + avx512vl_i16_info, avx512vl_i16_info>, VEX_W, EVEX_CD8<16, CD8VF>; +defm VPERMI2PS : avx512_perm_i_sizes<0x77, "vpermi2ps", + avx512vl_f32_info, avx512vl_i32_info>, EVEX_CD8<32, CD8VF>; +defm VPERMI2PD : avx512_perm_i_sizes<0x77, "vpermi2pd", + avx512vl_f64_info, avx512vl_i64_info>, VEX_W, EVEX_CD8<64, CD8VF>; + +// VPERMT2 +multiclass avx512_perm_t opc, string OpcodeStr, + X86VectorVTInfo _, X86VectorVTInfo IdxVT> { +let Constraints = "$src1 = $dst" in { + defm rr: AVX512_maskable_3src, EVEX_4V, + AVX5128IBase; + + let mayLoad = 1 in + defm rm: AVX512_maskable_3src, + EVEX_4V, AVX5128IBase; + } +} +multiclass avx512_perm_t_mb opc, string OpcodeStr, + X86VectorVTInfo _, X86VectorVTInfo IdxVT> { + let mayLoad = 1, Constraints = "$src1 = $dst" in + defm rmb: AVX512_maskable_3src, + AVX5128IBase, EVEX_4V, EVEX_B; +} + +multiclass avx512_perm_t_sizes opc, string OpcodeStr, + AVX512VLVectorVTInfo VTInfo, + AVX512VLVectorVTInfo ShuffleMask> { + defm NAME: avx512_perm_t, + avx512_perm_t_mb, EVEX_V512; let Predicates = [HasVLX] in { - defm NAME#128: avx512_perm_3src, - avx512_perm_3src_mb, - EVEX_V128; - defm NAME#256: avx512_perm_3src, - avx512_perm_3src_mb, - EVEX_V256; + defm NAME#128: avx512_perm_t, + avx512_perm_t_mb, EVEX_V128; + defm NAME#256: avx512_perm_t, + avx512_perm_t_mb, EVEX_V256; } } -multiclass avx512_perm_3src_sizes_w opc, string OpcodeStr, - SDNode OpNode, AVX512VLVectorVTInfo VTInfo> { + +multiclass avx512_perm_t_sizes_w opc, string OpcodeStr, + AVX512VLVectorVTInfo VTInfo, + AVX512VLVectorVTInfo Idx> { let Predicates = [HasBWI] in - defm NAME: avx512_perm_3src, - avx512_perm_3src_mb, - EVEX_V512; + defm NAME: avx512_perm_t, EVEX_V512; let Predicates = [HasBWI, HasVLX] in { - defm NAME#128: avx512_perm_3src, - avx512_perm_3src_mb, - EVEX_V128; - defm NAME#256: avx512_perm_3src, - avx512_perm_3src_mb, - 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_t, EVEX_V128; + defm NAME#256: avx512_perm_t, 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 @@ -2093,17 +2176,19 @@ let Predicates = [HasAVX512] in { (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. @@ -2313,16 +2398,6 @@ defm KUNPCKBW : avx512_mask_unpck<"bw", VK16, v16i1, VK8, HasAVX512>, PD; defm KUNPCKWD : avx512_mask_unpck<"wd", VK32, v32i1, VK16, HasBWI>, PS; defm KUNPCKDQ : avx512_mask_unpck<"dq", VK64, v64i1, VK32, HasBWI>, PS, VEX_W; -multiclass avx512_mask_unpck_int { - let Predicates = [HasAVX512] in - def : Pat<(!cast("int_x86_avx512_"##IntName##"_bw") - (i16 GR16:$src1), (i16 GR16:$src2)), - (COPY_TO_REGCLASS (!cast(InstName##"BWrr") - (v16i1 (COPY_TO_REGCLASS GR16:$src1, VK16)), - (v16i1 (COPY_TO_REGCLASS GR16:$src2, VK16))), GR16)>; -} -defm : avx512_mask_unpck_int<"kunpck", "KUNPCK">; - // Mask bit testing multiclass avx512_mask_testop opc, string OpcodeStr, RegisterClass KRC, SDNode OpNode, Predicate prd> { @@ -2413,24 +2488,52 @@ def : Pat<(v16i1 (insert_subvector undef, (v8i1 VK8:$src), (iPTR 0))), def : Pat<(v8i1 (extract_subvector (v16i1 VK16:$src), (iPTR 8))), (v8i1 (COPY_TO_REGCLASS (KSHIFTRWri VK16:$src, (i8 8)), VK8))>; +def : Pat<(v16i1 (extract_subvector (v32i1 VK32:$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:$src, VK32))>; def : Pat<(v32i1 (extract_subvector (v64i1 VK64:$src), (iPTR 32))), (v32i1 (COPY_TO_REGCLASS (KSHIFTRQri VK64:$src, (i8 32)), VK32))>; -let Predicates = [HasVLX] in { - def : Pat<(v8i1 (insert_subvector undef, (v4i1 VK4:$src), (iPTR 0))), - (v8i1 (COPY_TO_REGCLASS VK4:$src, VK8))>; - def : Pat<(v8i1 (insert_subvector undef, (v2i1 VK2:$src), (iPTR 0))), - (v8i1 (COPY_TO_REGCLASS VK2:$src, VK8))>; - def : Pat<(v4i1 (insert_subvector undef, (v2i1 VK2:$src), (iPTR 0))), - (v4i1 (COPY_TO_REGCLASS VK2:$src, VK4))>; - def : Pat<(v4i1 (extract_subvector (v8i1 VK8:$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:$src, VK2))>; -} +def : Pat<(v4i1 (extract_subvector (v8i1 VK8:$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:$src, VK2))>; + +def : Pat<(v4i1 (insert_subvector undef, (v2i1 VK2:$src), (iPTR 0))), + (v4i1 (COPY_TO_REGCLASS VK2:$src, VK4))>; + +def : Pat<(v8i1 (insert_subvector undef, (v4i1 VK4:$src), (iPTR 0))), + (v8i1 (COPY_TO_REGCLASS VK4:$src, VK8))>; +def : Pat<(v8i1 (insert_subvector undef, (v2i1 VK2:$src), (iPTR 0))), + (v8i1 (COPY_TO_REGCLASS VK2:$src, VK8))>; + +def : Pat<(v32i1 (insert_subvector undef, VK2:$src, (iPTR 0))), + (v32i1 (COPY_TO_REGCLASS VK2:$src, VK32))>; +def : Pat<(v32i1 (insert_subvector undef, VK4:$src, (iPTR 0))), + (v32i1 (COPY_TO_REGCLASS VK4:$src, VK32))>; +def : Pat<(v32i1 (insert_subvector undef, VK8:$src, (iPTR 0))), + (v32i1 (COPY_TO_REGCLASS VK8:$src, VK32))>; +def : Pat<(v32i1 (insert_subvector undef, VK16:$src, (iPTR 0))), + (v32i1 (COPY_TO_REGCLASS VK16:$src, VK32))>; + +def : Pat<(v64i1 (insert_subvector undef, VK2:$src, (iPTR 0))), + (v64i1 (COPY_TO_REGCLASS VK2:$src, VK64))>; +def : Pat<(v64i1 (insert_subvector undef, VK4:$src, (iPTR 0))), + (v64i1 (COPY_TO_REGCLASS VK4:$src, VK64))>; +def : Pat<(v64i1 (insert_subvector undef, VK8:$src, (iPTR 0))), + (v64i1 (COPY_TO_REGCLASS VK8:$src, VK64))>; +def : Pat<(v64i1 (insert_subvector undef, VK16:$src, (iPTR 0))), + (v64i1 (COPY_TO_REGCLASS VK16:$src, VK64))>; +def : Pat<(v64i1 (insert_subvector undef, VK32:$src, (iPTR 0))), + (v64i1 (COPY_TO_REGCLASS VK32:$src, VK64))>; + def : Pat<(v8i1 (X86vshli VK8:$src, (i8 imm:$imm))), (v8i1 (COPY_TO_REGCLASS @@ -2669,22 +2772,6 @@ def: Pat<(int_x86_avx512_mask_store_pd_512 addr:$ptr, (v8f64 VR512:$src), (VMOVAPDZmrk addr:$ptr, (v8i1 (COPY_TO_REGCLASS GR8:$mask, VK8WM)), VR512:$src)>; -let Predicates = [HasAVX512, NoVLX] in { -def: Pat<(X86mstore addr:$ptr, VK8WM:$mask, (v8f32 VR256:$src)), - (VMOVUPSZmrk addr:$ptr, - (v16i1 (COPY_TO_REGCLASS VK8WM:$mask, VK16WM)), - (INSERT_SUBREG (v16f32 (IMPLICIT_DEF)), VR256:$src, sub_ymm))>; - -def: Pat<(v8f32 (masked_load addr:$ptr, VK8WM:$mask, undef)), - (v8f32 (EXTRACT_SUBREG (v16f32 (VMOVUPSZrmkz - (v16i1 (COPY_TO_REGCLASS VK8WM:$mask, VK16WM)), addr:$ptr)), sub_ymm))>; - -def: Pat<(v8f32 (masked_load addr:$ptr, VK8WM:$mask, (v8f32 VR256:$src0))), - (v8f32 (EXTRACT_SUBREG (v16f32 (VMOVUPSZrmk - (INSERT_SUBREG (v16f32 (IMPLICIT_DEF)), VR256:$src0, sub_ymm), - (v16i1 (COPY_TO_REGCLASS VK8WM:$mask, VK16WM)), addr:$ptr)), sub_ymm))>; -} - defm VMOVDQA32 : avx512_alignedload_vl<0x6F, "vmovdqa32", avx512vl_i32_info, HasAVX512>, avx512_alignedstore_vl<0x7F, "vmovdqa32", avx512vl_i32_info, @@ -2746,17 +2833,6 @@ def : Pat<(v16i32 (vselect VK16WM:$mask, (v16i32 immAllZerosV), (v16i32 VR512:$src))), (VMOVDQU32Zrrkz (KNOTWrr VK16WM:$mask), VR512:$src)>; } -// NoVLX patterns -let Predicates = [HasAVX512, NoVLX] in { -def: Pat<(X86mstore addr:$ptr, VK8WM:$mask, (v8i32 VR256:$src)), - (VMOVDQU32Zmrk addr:$ptr, - (v16i1 (COPY_TO_REGCLASS VK8WM:$mask, VK16WM)), - (INSERT_SUBREG (v16i32 (IMPLICIT_DEF)), VR256:$src, sub_ymm))>; - -def: Pat<(v8i32 (masked_load addr:$ptr, VK8WM:$mask, undef)), - (v8i32 (EXTRACT_SUBREG (v16i32 (VMOVDQU32Zrmkz - (v16i1 (COPY_TO_REGCLASS VK8WM:$mask, VK16WM)), addr:$ptr)), sub_ymm))>; -} // Move Int Doubleword to Packed Double Int // @@ -2809,13 +2885,13 @@ def VMOVDI2SSZrm : AVX512BI<0x6E, MRMSrcMem, (outs FR32X:$dst), (ins i32mem:$sr // def VMOVPDI2DIZrr : AVX512BI<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128X:$src), "vmovd\t{$src, $dst|$dst, $src}", - [(set GR32:$dst, (vector_extract (v4i32 VR128X:$src), + [(set GR32:$dst, (extractelt (v4i32 VR128X:$src), (iPTR 0)))], IIC_SSE_MOVD_ToGP>, EVEX, VEX_LIG; def VMOVPDI2DIZmr : AVX512BI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR128X:$src), "vmovd\t{$src, $dst|$dst, $src}", - [(store (i32 (vector_extract (v4i32 VR128X:$src), + [(store (i32 (extractelt (v4i32 VR128X:$src), (iPTR 0))), addr:$dst)], IIC_SSE_MOVDQ>, EVEX, VEX_LIG, EVEX_CD8<32, CD8VT1>; @@ -2836,6 +2912,11 @@ def VMOVPQIto64Zmr : I<0xD6, MRMDestMem, (outs), EVEX, PD, VEX_LIG, VEX_W, EVEX_CD8<64, CD8VT1>, Sched<[WriteStore]>, Requires<[HasAVX512, In64BitMode]>; +def VMOV64toPQIZrr_REV : AVX512BI<0xD6, MRMDestReg, (outs VR128X:$dst), + (ins VR128X:$src), + "vmovq.s\t{$src, $dst|$dst, $src}",[]>, + EVEX, VEX_W, VEX_LIG; + // Move Scalar Single to Double Int // let isCodeGenOnly = 1 in { @@ -2864,53 +2945,60 @@ def VMOVQI2PQIZrm : AVX512BI<0x6E, MRMSrcMem, (outs VR128X:$dst), // AVX-512 MOVSS, MOVSD //===----------------------------------------------------------------------===// -multiclass avx512_move_scalar { - let hasSideEffects = 0 in { - def rr : SI<0x10, MRMSrcReg, (outs VR128X:$dst), (ins VR128X:$src1, RC:$src2), - !strconcat(asm, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"), - [(set VR128X:$dst, (vt (OpNode VR128X:$src1, - (scalar_to_vector RC:$src2))))], - IIC_SSE_MOV_S_RR>, EVEX_4V, VEX_LIG; - let Constraints = "$src1 = $dst" in - def rrk : SI<0x10, MRMSrcReg, (outs VR128X:$dst), - (ins VR128X:$src1, VK1WM:$mask, RC:$src2, RC:$src3), - !strconcat(asm, - "\t{$src3, $src2, $dst {${mask}}|$dst {${mask}}, $src2, $src3}"), - [], IIC_SSE_MOV_S_RR>, EVEX_4V, VEX_LIG, EVEX_K; - def rm : SI<0x10, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src), - !strconcat(asm, "\t{$src, $dst|$dst, $src}"), - [(set RC:$dst, (mem_pat addr:$src))], IIC_SSE_MOV_S_RM>, - EVEX, VEX_LIG; +multiclass avx512_move_scalar { + defm rr_Int : AVX512_maskable_scalar<0x10, MRMSrcReg, _, (outs _.RC:$dst), + (ins _.RC:$src1, _.RC:$src2), + asm, "$src2, $src1","$src1, $src2", + (_.VT (OpNode (_.VT _.RC:$src1), + (_.VT _.RC:$src2))), + IIC_SSE_MOV_S_RR>, EVEX_4V; + let Constraints = "$src1 = $dst" , mayLoad = 1 in + defm rm_Int : AVX512_maskable_3src_scalar<0x10, MRMSrcMem, _, + (outs _.RC:$dst), + (ins _.ScalarMemOp:$src), + asm,"$src","$src", + (_.VT (OpNode (_.VT _.RC:$src1), + (_.VT (scalar_to_vector + (_.ScalarLdFrag addr:$src)))))>, EVEX; + let isCodeGenOnly = 1 in { + def rr : AVX512PI<0x10, MRMSrcReg, (outs _.RC:$dst), + (ins _.RC:$src1, _.FRC:$src2), + !strconcat(asm, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"), + [(set _.RC:$dst, (_.VT (OpNode _.RC:$src1, + (scalar_to_vector _.FRC:$src2))))], + _.ExeDomain,IIC_SSE_MOV_S_RR>, EVEX_4V; + let mayLoad = 1 in + def rm : AVX512PI<0x10, MRMSrcMem, (outs _.FRC:$dst), (ins _.ScalarMemOp:$src), + !strconcat(asm, "\t{$src, $dst|$dst, $src}"), + [(set _.FRC:$dst, (_.ScalarLdFrag addr:$src))], + _.ExeDomain, IIC_SSE_MOV_S_RM>, EVEX; + } let mayStore = 1 in { - def mr: SI<0x11, MRMDestMem, (outs), (ins x86memop:$dst, RC:$src), - !strconcat(asm, "\t{$src, $dst|$dst, $src}"), - [(store RC:$src, addr:$dst)], IIC_SSE_MOV_S_MR>, - EVEX, VEX_LIG; - def mrk: SI<0x11, MRMDestMem, (outs), (ins x86memop:$dst, VK1WM:$mask, RC:$src), - !strconcat(asm, "\t{$src, $dst {${mask}}|$dst {${mask}}, $src}"), - [], IIC_SSE_MOV_S_MR>, - EVEX, VEX_LIG, EVEX_K; + def mr: AVX512PI<0x11, MRMDestMem, (outs), (ins _.ScalarMemOp:$dst, _.FRC:$src), + !strconcat(asm, "\t{$src, $dst|$dst, $src}"), + [(store _.FRC:$src, addr:$dst)], _.ExeDomain, IIC_SSE_MOV_S_MR>, + EVEX; + def mrk: AVX512PI<0x11, MRMDestMem, (outs), + (ins _.ScalarMemOp:$dst, VK1WM:$mask, _.FRC:$src), + !strconcat(asm, "\t{$src, $dst {${mask}}|$dst {${mask}}, $src}"), + [], _.ExeDomain, IIC_SSE_MOV_S_MR>, EVEX, EVEX_K; } // mayStore - } //hasSideEffects = 0 } -let ExeDomain = SSEPackedSingle in -defm VMOVSSZ : avx512_move_scalar<"movss", FR32X, X86Movss, v4f32, f32mem, - loadf32>, XS, EVEX_CD8<32, CD8VT1>; +defm VMOVSSZ : avx512_move_scalar<"vmovss", X86Movss, f32x_info>, + VEX_LIG, XS, EVEX_CD8<32, CD8VT1>; -let ExeDomain = SSEPackedDouble in -defm VMOVSDZ : avx512_move_scalar<"movsd", FR64X, X86Movsd, v2f64, f64mem, - loadf64>, XD, VEX_W, EVEX_CD8<64, CD8VT1>; +defm VMOVSDZ : avx512_move_scalar<"vmovsd", X86Movsd, f64x_info>, + VEX_LIG, XD, VEX_W, EVEX_CD8<64, CD8VT1>; def : Pat<(f32 (X86select VK1WM:$mask, (f32 FR32X:$src1), (f32 FR32X:$src2))), - (COPY_TO_REGCLASS (VMOVSSZrrk (COPY_TO_REGCLASS FR32X:$src2, VR128X), - VK1WM:$mask, (f32 (IMPLICIT_DEF)), FR32X:$src1), FR32X)>; + (COPY_TO_REGCLASS (VMOVSSZrr_Intk (COPY_TO_REGCLASS FR32X:$src2, VR128X), + VK1WM:$mask, (v4f32 (IMPLICIT_DEF)),(COPY_TO_REGCLASS FR32X:$src1, VR128X)), FR32X)>; def : Pat<(f64 (X86select VK1WM:$mask, (f64 FR64X:$src1), (f64 FR64X:$src2))), - (COPY_TO_REGCLASS (VMOVSDZrrk (COPY_TO_REGCLASS FR64X:$src2, VR128X), - VK1WM:$mask, (f64 (IMPLICIT_DEF)), FR64X:$src1), FR64X)>; + (COPY_TO_REGCLASS (VMOVSDZrr_Intk (COPY_TO_REGCLASS FR64X:$src2, VR128X), + VK1WM:$mask, (v2f64 (IMPLICIT_DEF)), (COPY_TO_REGCLASS FR64X:$src1, VR128X)), FR64X)>; def : Pat<(int_x86_avx512_mask_store_ss addr:$dst, VR128X:$src, GR8:$mask), (VMOVSSZmrk addr:$dst, (i1 (COPY_TO_REGCLASS GR8:$mask, VK1WM)), @@ -3008,10 +3096,10 @@ let Predicates = [HasAVX512] in { (EXTRACT_SUBREG (v4i64 VR256X:$src), sub_xmm)), sub_xmm)>; // Extract and store. - def : Pat<(store (f32 (vector_extract (v4f32 VR128X:$src), (iPTR 0))), + def : Pat<(store (f32 (extractelt (v4f32 VR128X:$src), (iPTR 0))), addr:$dst), (VMOVSSZmr addr:$dst, (COPY_TO_REGCLASS (v4f32 VR128X:$src), FR32X))>; - def : Pat<(store (f64 (vector_extract (v2f64 VR128X:$src), (iPTR 0))), + def : Pat<(store (f64 (extractelt (v2f64 VR128X:$src), (iPTR 0))), addr:$dst), (VMOVSDZmr addr:$dst, (COPY_TO_REGCLASS (v2f64 VR128X:$src), FR64X))>; @@ -4064,6 +4152,27 @@ multiclass avx512_var_shift_types opc, string OpcodeStr, avx512vl_i64_info>, VEX_W; } +// Use 512bit version to implement 128/256 bit in case NoVLX. +multiclass avx512_var_shift_w_lowering { + let Predicates = [HasBWI, NoVLX] in { + def : Pat<(_.info256.VT (OpNode (_.info256.VT _.info256.RC:$src1), + (_.info256.VT _.info256.RC:$src2))), + (EXTRACT_SUBREG + (!cast(NAME#"WZrr") + (INSERT_SUBREG (_.info512.VT (IMPLICIT_DEF)), VR256X:$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:$src2))), + (EXTRACT_SUBREG + (!cast(NAME#"WZrr") + (INSERT_SUBREG (_.info512.VT (IMPLICIT_DEF)), VR128X:$src1, sub_xmm), + (INSERT_SUBREG (_.info512.VT (IMPLICIT_DEF)), VR128X:$src2, sub_xmm)), + sub_xmm)>; + } +} + multiclass avx512_var_shift_w opc, string OpcodeStr, SDNode OpNode> { let Predicates = [HasBWI] in @@ -4079,11 +4188,14 @@ multiclass avx512_var_shift_w opc, string OpcodeStr, } defm VPSLLV : avx512_var_shift_types<0x47, "vpsllv", shl>, - avx512_var_shift_w<0x12, "vpsllvw", shl>; + avx512_var_shift_w<0x12, "vpsllvw", shl>, + avx512_var_shift_w_lowering; defm VPSRAV : avx512_var_shift_types<0x46, "vpsrav", sra>, - avx512_var_shift_w<0x11, "vpsravw", sra>; + avx512_var_shift_w<0x11, "vpsravw", sra>, + avx512_var_shift_w_lowering; defm VPSRLV : avx512_var_shift_types<0x45, "vpsrlv", srl>, - avx512_var_shift_w<0x10, "vpsrlvw", srl>; + avx512_var_shift_w<0x10, "vpsrlvw", srl>, + avx512_var_shift_w_lowering; defm VPRORV : avx512_var_shift_types<0x14, "vprorv", rotr>; defm VPROLV : avx512_var_shift_types<0x15, "vprolv", rotl>; @@ -5390,7 +5502,7 @@ defm VCVTQQ2PS : avx512_cvtqq2ps<0x5B, "vcvtqq2ps", sint_to_fp, defm VCVTUQQ2PS : avx512_cvtqq2ps<0x7A, "vcvtuqq2ps", uint_to_fp, X86VUlongToFpRnd>, VEX_W, XD, EVEX_CD8<64, CD8VF>; -let Predicates = [NoVLX] in { +let Predicates = [HasAVX512, NoVLX] in { def : Pat<(v8i32 (fp_to_uint (v8f32 VR256X:$src1))), (EXTRACT_SUBREG (v16i32 (VCVTTPS2UDQZrr (v16f32 (SUBREG_TO_REG (i32 0), VR256X:$src1, sub_ymm)))), sub_ymm)>; @@ -5496,6 +5608,29 @@ let Predicates = [HasAVX512] in { EVEX, EVEX_V128, EVEX_CD8<32, CD8VH>; } } + +// Unordered/Ordered scalar fp compare with Sea and set EFLAGS +multiclass avx512_ord_cmp_sae opc, X86VectorVTInfo _, SDNode OpNode, + string OpcodeStr> { + def rb: AVX512, EVEX, EVEX_B, VEX_LIG, EVEX_V128, + Sched<[WriteFAdd]>; +} + +let Defs = [EFLAGS], Predicates = [HasAVX512] in { + defm VUCOMISSZ : avx512_ord_cmp_sae<0x2E, v4f32x_info, X86ucomiSae, "vucomiss">, + AVX512PSIi8Base, EVEX_CD8<32, CD8VT1>; + defm VUCOMISDZ : avx512_ord_cmp_sae<0x2E, v2f64x_info, X86ucomiSae, "vucomisd">, + AVX512PDIi8Base, VEX_W, EVEX_CD8<64, CD8VT1>; + defm VCOMISSZ : avx512_ord_cmp_sae<0x2F, v4f32x_info, X86comiSae, "vcomiss">, + AVX512PSIi8Base, EVEX_CD8<32, CD8VT1>; + defm VCOMISDZ : avx512_ord_cmp_sae<0x2F, v2f64x_info, X86comiSae, "vcomisd">, + AVX512PDIi8Base, VEX_W, EVEX_CD8<64, CD8VT1>; +} + let Defs = [EFLAGS], Predicates = [HasAVX512] in { defm VUCOMISSZ : sse12_ord_cmp<0x2E, FR32X, X86cmp, f32, f32mem, loadf32, "ucomiss">, PS, EVEX, VEX_LIG, @@ -5601,20 +5736,6 @@ multiclass avx512_fp14_p_vl_all opc, string OpcodeStr, SDNode OpNode> { defm VRSQRT14 : avx512_fp14_p_vl_all<0x4E, "vrsqrt14", X86frsqrt>; defm VRCP14 : avx512_fp14_p_vl_all<0x4C, "vrcp14", X86frcp>; -def : Pat <(v16f32 (int_x86_avx512_rsqrt14_ps_512 (v16f32 VR512:$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:$src)>; - /// avx512_fp28_s rcp28ss, rcp28sd, rsqrt28ss, rsqrt28sd multiclass avx512_fp28_s opc, string OpcodeStr,X86VectorVTInfo _, SDNode OpNode> { @@ -5802,12 +5923,12 @@ multiclass avx512_sqrt_scalar opc, string OpcodeStr,X86VectorVTInfo _, EVEX_B, EVEX_RC; let isCodeGenOnly = 1 in { - def r : SI; let mayLoad = 1 in - def m : SI; } @@ -7348,32 +7469,34 @@ defm VPSRLDQ : avx512_shift_packed_all<0x73, X86vshrdq, MRM3r, MRM3m, "vpsrldq", multiclass avx512_psadbw_packed opc, SDNode OpNode, - string OpcodeStr, X86VectorVTInfo _src>{ + string OpcodeStr, X86VectorVTInfo _dst, + X86VectorVTInfo _src>{ def rr : AVX512BI; + [(set _dst.RC:$dst,(_dst.VT + (OpNode (_src.VT _src.RC:$src1), + (_src.VT _src.RC:$src2))))]>; let mayLoad = 1 in def rm : AVX512BI; } multiclass avx512_psadbw_packed_all opc, SDNode OpNode, string OpcodeStr, Predicate prd> { let Predicates = [prd] in - defm Z512 : avx512_psadbw_packed, - EVEX_V512; + defm Z512 : avx512_psadbw_packed, EVEX_V512; let Predicates = [prd, HasVLX] in { - defm Z256 : avx512_psadbw_packed, - EVEX_V256; - defm Z128 : avx512_psadbw_packed, - EVEX_V128; + defm Z256 : avx512_psadbw_packed, EVEX_V256; + defm Z128 : avx512_psadbw_packed, EVEX_V128; } }