diff options
Diffstat (limited to 'lib/Target/X86/X86InstrAVX512.td')
| -rw-r--r-- | lib/Target/X86/X86InstrAVX512.td | 446 |
1 files changed, 329 insertions, 117 deletions
diff --git a/lib/Target/X86/X86InstrAVX512.td b/lib/Target/X86/X86InstrAVX512.td index 9d11d3c..c1d0aef 100644 --- a/lib/Target/X86/X86InstrAVX512.td +++ b/lib/Target/X86/X86InstrAVX512.td @@ -1047,12 +1047,6 @@ multiclass avx512_permil<bits<8> OpcImm, bits<8> OpcVar, X86VectorVTInfo _, EVEX_4V; } } - -defm VPERMQZ : avx512_perm_imm<0x00, "vpermq", X86VPermi, v8i64_info>, - EVEX_V512, VEX_W; -defm VPERMPDZ : avx512_perm_imm<0x01, "vpermpd", X86VPermi, v8f64_info>, - EVEX_V512, VEX_W; - defm VPERMILPSZ : avx512_permil<0x04, 0x0C, v16f32_info, v16i32_info>, EVEX_V512; defm VPERMILPDZ : avx512_permil<0x05, 0x0D, v8f64_info, v8i64_info>, @@ -1063,37 +1057,6 @@ def : Pat<(v16i32 (X86VPermilpi VR512:$src1, (i8 imm:$imm))), def : Pat<(v8i64 (X86VPermilpi VR512:$src1, (i8 imm:$imm))), (VPERMILPDZri VR512:$src1, imm:$imm)>; -// -- VPERM - register form -- -multiclass avx512_perm<bits<8> opc, string OpcodeStr, RegisterClass RC, - PatFrag mem_frag, X86MemOperand x86memop, ValueType OpVT> { - - def rr : AVX5128I<opc, MRMSrcReg, (outs RC:$dst), - (ins RC:$src1, RC:$src2), - !strconcat(OpcodeStr, - "\t{$src2, $src1, $dst|$dst, $src1, $src2}"), - [(set RC:$dst, - (OpVT (X86VPermv RC:$src1, RC:$src2)))]>, EVEX_4V; - - def rm : AVX5128I<opc, MRMSrcMem, (outs RC:$dst), - (ins RC:$src1, x86memop:$src2), - !strconcat(OpcodeStr, - "\t{$src2, $src1, $dst|$dst, $src1, $src2}"), - [(set RC:$dst, - (OpVT (X86VPermv RC:$src1, (mem_frag addr:$src2))))]>, - EVEX_4V; -} - -defm VPERMDZ : avx512_perm<0x36, "vpermd", VR512, loadv16i32, i512mem, - v16i32>, EVEX_V512, EVEX_CD8<32, CD8VF>; -defm VPERMQZ : avx512_perm<0x36, "vpermq", VR512, loadv8i64, i512mem, - v8i64>, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>; -let ExeDomain = SSEPackedSingle in -defm VPERMPSZ : avx512_perm<0x16, "vpermps", VR512, loadv16f32, f512mem, - v16f32>, EVEX_V512, EVEX_CD8<32, CD8VF>; -let ExeDomain = SSEPackedDouble in -defm VPERMPDZ : avx512_perm<0x16, "vpermpd", VR512, loadv8f64, f512mem, - v8f64>, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>; - // -- VPERM2I - 3 source operands form -- multiclass avx512_perm_3src<bits<8> opc, string OpcodeStr, RegisterClass RC, PatFrag mem_frag, X86MemOperand x86memop, @@ -3401,32 +3364,6 @@ defm VPUNPCKHQDQZ : avx512_unpack_int<0x6D, "vpunpckhqdq", X86Unpckh, v8i64, VR512, loadv8i64, i512mem>, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>; //===----------------------------------------------------------------------===// -// AVX-512 - PSHUFD -// - -multiclass avx512_pshuf_imm<bits<8> opc, string OpcodeStr, RegisterClass RC, - SDNode OpNode, PatFrag mem_frag, - X86MemOperand x86memop, ValueType OpVT> { - def ri : AVX512Ii8<opc, MRMSrcReg, (outs RC:$dst), - (ins RC:$src1, u8imm:$src2), - !strconcat(OpcodeStr, - "\t{$src2, $src1, $dst|$dst, $src1, $src2}"), - [(set RC:$dst, - (OpVT (OpNode RC:$src1, (i8 imm:$src2))))]>, - EVEX; - def mi : AVX512Ii8<opc, MRMSrcMem, (outs RC:$dst), - (ins x86memop:$src1, u8imm:$src2), - !strconcat(OpcodeStr, - "\t{$src2, $src1, $dst|$dst, $src1, $src2}"), - [(set RC:$dst, - (OpVT (OpNode (mem_frag addr:$src1), - (i8 imm:$src2))))]>, EVEX; -} - -defm VPSHUFDZ : avx512_pshuf_imm<0x70, "vpshufd", VR512, X86PShufd, loadv16i32, - i512mem, v16i32>, PD, EVEX_V512, EVEX_CD8<32, CD8VF>; - -//===----------------------------------------------------------------------===// // AVX-512 Logical Instructions //===----------------------------------------------------------------------===// @@ -3729,14 +3666,14 @@ multiclass avx512_shift_rmi<bits<8> opc, Format ImmFormR, Format ImmFormM, (ins _.RC:$src1, u8imm:$src2), OpcodeStr, "$src2, $src1", "$src1, $src2", (_.VT (OpNode _.RC:$src1, (i8 imm:$src2))), - SSE_INTSHIFT_ITINS_P.rr>, AVX512BIi8Base, EVEX_4V; + SSE_INTSHIFT_ITINS_P.rr>; let mayLoad = 1 in defm mi : AVX512_maskable<opc, ImmFormM, _, (outs _.RC:$dst), (ins _.MemOp:$src1, u8imm:$src2), OpcodeStr, "$src2, $src1", "$src1, $src2", (_.VT (OpNode (_.VT (bitconvert (_.LdFrag addr:$src1))), (i8 imm:$src2))), - SSE_INTSHIFT_ITINS_P.rm>, AVX512BIi8Base, EVEX_4V; + SSE_INTSHIFT_ITINS_P.rm>; } multiclass avx512_shift_rmbi<bits<8> opc, Format ImmFormM, @@ -3746,7 +3683,7 @@ multiclass avx512_shift_rmbi<bits<8> opc, Format ImmFormM, (ins _.ScalarMemOp:$src1, u8imm:$src2), OpcodeStr, "$src2, ${src1}"##_.BroadcastStr, "${src1}"##_.BroadcastStr##", $src2", (_.VT (OpNode (X86VBroadcast (_.ScalarLdFrag addr:$src1)), (i8 imm:$src2))), - SSE_INTSHIFT_ITINS_P.rm>, AVX512BIi8Base, EVEX_4V, EVEX_B; + SSE_INTSHIFT_ITINS_P.rm>, EVEX_B; } multiclass avx512_shift_rrm<bits<8> opc, string OpcodeStr, SDNode OpNode, @@ -3836,16 +3773,16 @@ multiclass avx512_shift_rmi_dq<bits<8> opcd, bits<8> opcq, } defm VPSRL : avx512_shift_rmi_dq<0x72, 0x73, MRM2r, MRM2m, "vpsrl", X86vsrli>, - avx512_shift_rmi_w<0x71, MRM2r, MRM2m, "vpsrlw", X86vsrli>; + avx512_shift_rmi_w<0x71, MRM2r, MRM2m, "vpsrlw", X86vsrli>, AVX512BIi8Base, EVEX_4V; defm VPSLL : avx512_shift_rmi_dq<0x72, 0x73, MRM6r, MRM6m, "vpsll", X86vshli>, - avx512_shift_rmi_w<0x71, MRM6r, MRM6m, "vpsllw", X86vshli>; + avx512_shift_rmi_w<0x71, MRM6r, MRM6m, "vpsllw", X86vshli>, AVX512BIi8Base, EVEX_4V; defm VPSRA : avx512_shift_rmi_dq<0x72, 0x72, MRM4r, MRM4m, "vpsra", X86vsrai>, - avx512_shift_rmi_w<0x71, MRM4r, MRM4m, "vpsraw", X86vsrai>; + avx512_shift_rmi_w<0x71, MRM4r, MRM4m, "vpsraw", X86vsrai>, AVX512BIi8Base, EVEX_4V; -defm VPROR : avx512_shift_rmi_dq<0x72, 0x72, MRM0r, MRM0m, "vpror", rotr>; -defm VPROL : avx512_shift_rmi_dq<0x72, 0x72, MRM1r, MRM1m, "vprol", rotl>; +defm VPROR : avx512_shift_rmi_dq<0x72, 0x72, MRM0r, MRM0m, "vpror", rotr>, AVX512BIi8Base, EVEX_4V; +defm VPROL : avx512_shift_rmi_dq<0x72, 0x72, MRM1r, MRM1m, "vprol", rotl>, AVX512BIi8Base, EVEX_4V; defm VPSLL : avx512_shift_types<0xF2, 0xF3, 0xF1, "vpsll", X86vshl>; defm VPSRA : avx512_shift_types<0xE2, 0xE2, 0xE1, "vpsra", X86vsra>; @@ -3865,7 +3802,8 @@ multiclass avx512_var_shift<bits<8> opc, string OpcodeStr, SDNode OpNode, defm rm : AVX512_maskable<opc, MRMSrcMem, _, (outs _.RC:$dst), (ins _.RC:$src1, _.MemOp:$src2), OpcodeStr, "$src2, $src1", "$src1, $src2", - (_.VT (OpNode _.RC:$src1, (_.LdFrag addr:$src2))), + (_.VT (OpNode _.RC:$src1, + (_.VT (bitconvert (_.LdFrag addr:$src2))))), SSE_INTSHIFT_ITINS_P.rm>, AVX5128IBase, EVEX_4V, EVEX_CD8<_.EltSize, CD8VF>; } @@ -3927,6 +3865,65 @@ defm VPSRLV : avx512_var_shift_types<0x45, "vpsrlv", srl>, defm VPRORV : avx512_var_shift_types<0x14, "vprorv", rotr>; defm VPROLV : avx512_var_shift_types<0x15, "vprolv", rotl>; +//===-------------------------------------------------------------------===// +// 1-src variable permutation VPERMW/D/Q +//===-------------------------------------------------------------------===// +multiclass avx512_vperm_dq_sizes<bits<8> opc, string OpcodeStr, SDNode OpNode, + AVX512VLVectorVTInfo _> { + let Predicates = [HasAVX512] in + defm Z : avx512_var_shift<opc, OpcodeStr, OpNode, _.info512>, + avx512_var_shift_mb<opc, OpcodeStr, OpNode, _.info512>, EVEX_V512; + + let Predicates = [HasAVX512, HasVLX] in + defm Z256 : avx512_var_shift<opc, OpcodeStr, OpNode, _.info256>, + avx512_var_shift_mb<opc, OpcodeStr, OpNode, _.info256>, EVEX_V256; +} + +multiclass avx512_vpermi_dq_sizes<bits<8> opc, Format ImmFormR, Format ImmFormM, + string OpcodeStr, SDNode OpNode, + AVX512VLVectorVTInfo VTInfo> { + let Predicates = [HasAVX512] in + defm Z: avx512_shift_rmi<opc, ImmFormR, ImmFormM, OpcodeStr, OpNode, + VTInfo.info512>, + avx512_shift_rmbi<opc, ImmFormM, OpcodeStr, OpNode, + VTInfo.info512>, EVEX_V512; + let Predicates = [HasAVX512, HasVLX] in + defm Z256: avx512_shift_rmi<opc, ImmFormR, ImmFormM, OpcodeStr, OpNode, + VTInfo.info256>, + avx512_shift_rmbi<opc, ImmFormM, OpcodeStr, OpNode, + VTInfo.info256>, EVEX_V256; +} + + +defm VPERM : avx512_var_shift_w<0x8D, "vpermw", X86VPermv>; + +defm VPERMD : avx512_vperm_dq_sizes<0x36, "vpermd", X86VPermv, + avx512vl_i32_info>; +defm VPERMQ : avx512_vperm_dq_sizes<0x36, "vpermq", X86VPermv, + avx512vl_i64_info>, VEX_W; +defm VPERMPS : avx512_vperm_dq_sizes<0x16, "vpermps", X86VPermv, + avx512vl_f32_info>; +defm VPERMPD : avx512_vperm_dq_sizes<0x16, "vpermpd", X86VPermv, + avx512vl_f64_info>, VEX_W; + +defm VPERMQ : avx512_vpermi_dq_sizes<0x00, MRMSrcReg, MRMSrcMem, "vpermq", + X86VPermi, avx512vl_i64_info>, + EVEX, AVX512AIi8Base, EVEX_CD8<64, CD8VF>, VEX_W; +defm VPERMPD : avx512_vpermi_dq_sizes<0x01, MRMSrcReg, MRMSrcMem, "vpermpd", + X86VPermi, avx512vl_f64_info>, + EVEX, AVX512AIi8Base, EVEX_CD8<64, CD8VF>, VEX_W; + +//===----------------------------------------------------------------------===// +// AVX-512 - VPSHUFD, VPSHUFLW, VPSHUFHW +//===----------------------------------------------------------------------===// + +defm VPSHUFD : avx512_shift_rmi_sizes<0x70, MRMSrcReg, MRMSrcMem, "vpshufd", + X86PShufd, avx512vl_i32_info>, + EVEX, AVX512BIi8Base, EVEX_CD8<32, CD8VF>; +defm VPSHUFH : avx512_shift_rmi_w<0x70, MRMSrcReg, MRMSrcMem, "vpshufhw", + X86PShufhw>, EVEX, AVX512XSIi8Base, VEX_W; +defm VPSHUFL : avx512_shift_rmi_w<0x70, MRMSrcReg, MRMSrcMem, "vpshuflw", + X86PShuflw>, EVEX, AVX512XDIi8Base, VEX_W; //===----------------------------------------------------------------------===// // AVX-512 - MOVDDUP //===----------------------------------------------------------------------===// @@ -4869,11 +4866,6 @@ multiclass avx512_fp28_p<bits<8> opc, string OpcodeStr, X86VectorVTInfo _, (ins _.RC:$src), OpcodeStr, "$src", "$src", (OpNode (_.VT _.RC:$src), (i32 FROUND_CURRENT))>; - defm rb : AVX512_maskable<opc, MRMSrcReg, _, (outs _.RC:$dst), - (ins _.RC:$src), OpcodeStr, - "{sae}, $src", "$src, {sae}", - (OpNode (_.VT _.RC:$src), (i32 FROUND_NO_EXC))>, EVEX_B; - defm m : AVX512_maskable<opc, MRMSrcMem, _, (outs _.RC:$dst), (ins _.MemOp:$src), OpcodeStr, "$src", "$src", (OpNode (_.FloatVT @@ -4881,24 +4873,58 @@ multiclass avx512_fp28_p<bits<8> opc, string OpcodeStr, X86VectorVTInfo _, (i32 FROUND_CURRENT))>; defm mb : AVX512_maskable<opc, MRMSrcMem, _, (outs _.RC:$dst), - (ins _.MemOp:$src), OpcodeStr, "$src", "$src", + (ins _.MemOp:$src), OpcodeStr, + "${src}"##_.BroadcastStr, "${src}"##_.BroadcastStr, (OpNode (_.FloatVT (X86VBroadcast (_.ScalarLdFrag addr:$src))), (i32 FROUND_CURRENT))>, EVEX_B; } +multiclass avx512_fp28_p_round<bits<8> opc, string OpcodeStr, X86VectorVTInfo _, + SDNode OpNode> { + defm rb : AVX512_maskable<opc, MRMSrcReg, _, (outs _.RC:$dst), + (ins _.RC:$src), OpcodeStr, + "{sae}, $src", "$src, {sae}", + (OpNode (_.VT _.RC:$src), (i32 FROUND_NO_EXC))>, EVEX_B; +} multiclass avx512_eri<bits<8> opc, string OpcodeStr, SDNode OpNode> { defm PS : avx512_fp28_p<opc, OpcodeStr#"ps", v16f32_info, OpNode>, - EVEX_CD8<32, CD8VF>; + avx512_fp28_p_round<opc, OpcodeStr#"ps", v16f32_info, OpNode>, + T8PD, EVEX_V512, EVEX_CD8<32, CD8VF>; defm PD : avx512_fp28_p<opc, OpcodeStr#"pd", v8f64_info, OpNode>, - VEX_W, EVEX_CD8<32, CD8VF>; + avx512_fp28_p_round<opc, OpcodeStr#"pd", v8f64_info, OpNode>, + T8PD, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>; } +multiclass avx512_fp_unaryop_packed<bits<8> opc, string OpcodeStr, + SDNode OpNode> { + // Define only if AVX512VL feature is present. + let Predicates = [HasVLX] in { + defm PSZ128 : avx512_fp28_p<opc, OpcodeStr#"ps", v4f32x_info, OpNode>, + EVEX_V128, T8PD, EVEX_CD8<32, CD8VF>; + defm PSZ256 : avx512_fp28_p<opc, OpcodeStr#"ps", v8f32x_info, OpNode>, + EVEX_V256, T8PD, EVEX_CD8<32, CD8VF>; + defm PDZ128 : avx512_fp28_p<opc, OpcodeStr#"pd", v2f64x_info, OpNode>, + EVEX_V128, VEX_W, T8PD, EVEX_CD8<64, CD8VF>; + defm PDZ256 : avx512_fp28_p<opc, OpcodeStr#"pd", v4f64x_info, OpNode>, + EVEX_V256, VEX_W, T8PD, EVEX_CD8<64, CD8VF>; + } +} let Predicates = [HasERI], hasSideEffects = 0 in { - defm VRSQRT28 : avx512_eri<0xCC, "vrsqrt28", X86rsqrt28>, EVEX, EVEX_V512, T8PD; - defm VRCP28 : avx512_eri<0xCA, "vrcp28", X86rcp28>, EVEX, EVEX_V512, T8PD; - defm VEXP2 : avx512_eri<0xC8, "vexp2", X86exp2>, EVEX, EVEX_V512, T8PD; + defm VRSQRT28 : avx512_eri<0xCC, "vrsqrt28", X86rsqrt28>, EVEX; + defm VRCP28 : avx512_eri<0xCA, "vrcp28", X86rcp28>, EVEX; + defm VEXP2 : avx512_eri<0xC8, "vexp2", X86exp2>, EVEX; +} +defm VGETEXP : avx512_eri<0x42, "vgetexp", X86fgetexpRnd>, + avx512_fp_unaryop_packed<0x42, "vgetexp", X86fgetexpRnd> , EVEX; + +multiclass avx512_sqrt_packed_round<bits<8> opc, string OpcodeStr, + SDNode OpNodeRnd, X86VectorVTInfo _>{ + defm rb: AVX512_maskable<opc, MRMSrcReg, _, (outs _.RC:$dst), + (ins _.RC:$src, AVX512RC:$rc), OpcodeStr, "$rc, $src", "$src, $rc", + (_.VT (OpNodeRnd _.RC:$src, (i32 imm:$rc)))>, + EVEX, EVEX_B, EVEX_RC; } multiclass avx512_sqrt_packed<bits<8> opc, string OpcodeStr, @@ -5007,20 +5033,22 @@ multiclass avx512_sqrt_packed_all<bits<8> opc, string OpcodeStr, } } -defm VSQRT : avx512_sqrt_packed_all<0x51, "vsqrt", fsqrt>; +multiclass avx512_sqrt_packed_all_round<bits<8> opc, string OpcodeStr, + SDNode OpNodeRnd> { + defm PSZ : avx512_sqrt_packed_round<opc, !strconcat(OpcodeStr, "ps"), OpNodeRnd, + v16f32_info>, EVEX_V512, PS, EVEX_CD8<32, CD8VF>; + defm PDZ : avx512_sqrt_packed_round<opc, !strconcat(OpcodeStr, "pd"), OpNodeRnd, + v8f64_info>, EVEX_V512, VEX_W, PD, EVEX_CD8<64, CD8VF>; +} + +defm VSQRT : avx512_sqrt_packed_all<0x51, "vsqrt", fsqrt>, + avx512_sqrt_packed_all_round<0x51, "vsqrt", X86fsqrtRnd>; defm VSQRT : avx512_sqrt_scalar<0x51, "sqrt", int_x86_avx512_sqrt_ss, int_x86_avx512_sqrt_sd, SSE_SQRTSS, SSE_SQRTSD>; let Predicates = [HasAVX512] in { - def : Pat<(v16f32 (int_x86_avx512_sqrt_ps_512 (v16f32 VR512:$src1), - (bc_v16f32 (v16i32 immAllZerosV)), (i16 -1), FROUND_CURRENT)), - (VSQRTPSZr VR512:$src1)>; - def : Pat<(v8f64 (int_x86_avx512_sqrt_pd_512 (v8f64 VR512:$src1), - (bc_v8f64 (v16i32 immAllZerosV)), (i8 -1), FROUND_CURRENT)), - (VSQRTPDZr VR512:$src1)>; - def : Pat<(f32 (fsqrt FR32X:$src)), (VSQRTSSZr (f32 (IMPLICIT_DEF)), FR32X:$src)>; def : Pat<(f32 (fsqrt (load addr:$src))), @@ -5583,30 +5611,6 @@ def : Pat<(v8i64 (X86Shufp VR512:$src1, (loadv8i64 addr:$src2), (i8 imm:$imm))), (VSHUFPDZrmi VR512:$src1, addr:$src2, imm:$imm)>; -multiclass avx512_valign<X86VectorVTInfo _> { - defm rri : AVX512_maskable<0x03, MRMSrcReg, _, (outs _.RC:$dst), - (ins _.RC:$src1, _.RC:$src2, u8imm:$src3), - "valign"##_.Suffix, - "$src3, $src2, $src1", "$src1, $src2, $src3", - (_.VT (X86VAlign _.RC:$src2, _.RC:$src1, - (i8 imm:$src3)))>, - AVX512AIi8Base, EVEX_4V; - - // Also match valign of packed floats. - def : Pat<(_.FloatVT (X86VAlign _.RC:$src1, _.RC:$src2, (i8 imm:$imm))), - (!cast<Instruction>(NAME##rri) _.RC:$src2, _.RC:$src1, imm:$imm)>; - - let mayLoad = 1 in - def rmi : AVX512AIi8<0x03, MRMSrcMem, (outs _.RC:$dst), - (ins _.RC:$src1, _.MemOp:$src2, u8imm:$src3), - !strconcat("valign"##_.Suffix, - "\t{$src3, $src2, $src1, $dst|" - "$dst, $src1, $src2, $src3}"), - []>, EVEX_4V; -} -defm VALIGND : avx512_valign<v16i32_info>, EVEX_V512, EVEX_CD8<32, CD8VF>; -defm VALIGNQ : avx512_valign<v8i64_info>, VEX_W, EVEX_V512, EVEX_CD8<64, CD8VF>; - // Helper fragments to match sext vXi1 to vXiY. def v16i1sextv16i32 : PatLeaf<(v16i32 (X86vsrai VR512:$src, (i8 31)))>; def v8i1sextv8i64 : PatLeaf<(v8i64 (X86vsrai VR512:$src, (i8 63)))>; @@ -5949,7 +5953,7 @@ multiclass expand_by_vec_width<bits<8> opc, X86VectorVTInfo _, (_.LdFrag addr:$src))), _.RC:$src0)))]>, EVEX_K, EVEX_CD8<_.EltSize, CD8VT1>; - + let mayLoad = 1 in def rmkz : AVX5128I<opc, MRMSrcMem, (outs _.RC:$dst), (ins _.KRCWM:$mask, _.MemOp:$src), @@ -5958,7 +5962,6 @@ multiclass expand_by_vec_width<bits<8> opc, X86VectorVTInfo _, (_.VT (bitconvert (_.LdFrag addr:$src))), _.ImmAllZerosV)))]>, EVEX_KZ, EVEX_CD8<_.EltSize, CD8VT1>; - } multiclass expand_by_elt_width<bits<8> opc, string OpcodeStr, @@ -5979,3 +5982,212 @@ defm VEXPANDPS : expand_by_elt_width <0x88, "vexpandps", avx512vl_f32_info>, EVEX; defm VEXPANDPD : expand_by_elt_width <0x88, "vexpandpd", avx512vl_f64_info>, EVEX, VEX_W; + +//handle instruction reg_vec1 = op(reg_vec2,reg_vec3,imm) +// op(reg_vec2,mem_vec,imm) +// op(reg_vec2,broadcast(eltVt),imm) +//all instruction created with FROUND_CURRENT +multiclass avx512_fp_packed_imm<bits<8> opc, string OpcodeStr, SDNode OpNode, + X86VectorVTInfo _>{ + defm rri : AVX512_maskable<opc, MRMSrcReg, _, (outs _.RC:$dst), + (ins _.RC:$src1, _.RC:$src2, u8imm:$src3), + OpcodeStr, "$src3, $src2, $src1", "$src1, $src2, $src3", + (OpNode (_.VT _.RC:$src1), + (_.VT _.RC:$src2), + (i8 imm:$src3), + (i32 FROUND_CURRENT))>; + let mayLoad = 1 in { + defm rmi : AVX512_maskable<opc, MRMSrcMem, _, (outs _.RC:$dst), + (ins _.RC:$src1, _.MemOp:$src2, u8imm:$src3), + OpcodeStr, "$src3, $src2, $src1", "$src1, $src2, $src3", + (OpNode (_.VT _.RC:$src1), + (_.VT (bitconvert (_.LdFrag addr:$src2))), + (i8 imm:$src3), + (i32 FROUND_CURRENT))>; + defm rmbi : AVX512_maskable<opc, MRMSrcMem, _, (outs _.RC:$dst), + (ins _.RC:$src1, _.ScalarMemOp:$src2, u8imm:$src3), + OpcodeStr, "$src3, ${src2}"##_.BroadcastStr##", $src1", + "$src1, ${src2}"##_.BroadcastStr##", $src3", + (OpNode (_.VT _.RC:$src1), + (_.VT (X86VBroadcast(_.ScalarLdFrag addr:$src2))), + (i8 imm:$src3), + (i32 FROUND_CURRENT))>, EVEX_B; + } +} + +//handle instruction reg_vec1 = op(reg_vec2,reg_vec3,imm) +// op(reg_vec2,mem_vec,imm) +// op(reg_vec2,broadcast(eltVt),imm) +multiclass avx512_3Op_imm8<bits<8> opc, string OpcodeStr, SDNode OpNode, + X86VectorVTInfo _>{ + defm rri : AVX512_maskable<opc, MRMSrcReg, _, (outs _.RC:$dst), + (ins _.RC:$src1, _.RC:$src2, u8imm:$src3), + OpcodeStr, "$src3, $src2, $src1", "$src1, $src2, $src3", + (OpNode (_.VT _.RC:$src1), + (_.VT _.RC:$src2), + (i8 imm:$src3))>; + let mayLoad = 1 in { + defm rmi : AVX512_maskable<opc, MRMSrcMem, _, (outs _.RC:$dst), + (ins _.RC:$src1, _.MemOp:$src2, u8imm:$src3), + OpcodeStr, "$src3, $src2, $src1", "$src1, $src2, $src3", + (OpNode (_.VT _.RC:$src1), + (_.VT (bitconvert (_.LdFrag addr:$src2))), + (i8 imm:$src3))>; + defm rmbi : AVX512_maskable<opc, MRMSrcMem, _, (outs _.RC:$dst), + (ins _.RC:$src1, _.ScalarMemOp:$src2, u8imm:$src3), + OpcodeStr, "$src3, ${src2}"##_.BroadcastStr##", $src1", + "$src1, ${src2}"##_.BroadcastStr##", $src3", + (OpNode (_.VT _.RC:$src1), + (_.VT (X86VBroadcast(_.ScalarLdFrag addr:$src2))), + (i8 imm:$src3))>, EVEX_B; + } +} + +//handle scalar instruction reg_vec1 = op(reg_vec2,reg_vec3,imm) +// op(reg_vec2,mem_scalar,imm) +//all instruction created with FROUND_CURRENT +multiclass avx512_fp_scalar_imm<bits<8> opc, string OpcodeStr, SDNode OpNode, + X86VectorVTInfo _> { + + defm rri : AVX512_maskable_scalar<opc, MRMSrcReg, _, (outs _.RC:$dst), + (ins _.RC:$src1, _.RC:$src2, u8imm:$src3), + OpcodeStr, "$src3, $src2, $src1", "$src1, $src2, $src3", + (OpNode (_.VT _.RC:$src1), + (_.VT _.RC:$src2), + (i8 imm:$src3), + (i32 FROUND_CURRENT))>; + let mayLoad = 1 in { + defm rmi : AVX512_maskable_scalar<opc, MRMSrcMem, _, (outs _.RC:$dst), + (ins _.RC:$src1, _.MemOp:$src2, u8imm:$src3), + OpcodeStr, "$src3, $src2, $src1", "$src1, $src2, $src3", + (OpNode (_.VT _.RC:$src1), + (_.VT (scalar_to_vector + (_.ScalarLdFrag addr:$src2))), + (i8 imm:$src3), + (i32 FROUND_CURRENT))>; + + let isAsmParserOnly = 1 in { + defm rmi_alt :AVX512_maskable_in_asm<opc, MRMSrcMem, _, (outs _.FRC:$dst), + (ins _.FRC:$src1, _.ScalarMemOp:$src2, u8imm:$src3), + OpcodeStr, "$src3, $src2, $src1", "$src1, $src2, $src3", + []>; + } + } +} + +//handle instruction reg_vec1 = op(reg_vec2,reg_vec3,imm),{sae} +multiclass avx512_fp_sae_packed_imm<bits<8> opc, string OpcodeStr, + SDNode OpNode, X86VectorVTInfo _>{ + defm rrib : AVX512_maskable<opc, MRMSrcReg, _, (outs _.RC:$dst), + (ins _.RC:$src1, _.RC:$src2, u8imm:$src3), + OpcodeStr, "$src3,{sae}, $src2, $src1", + "$src1, $src2,{sae}, $src3", + (OpNode (_.VT _.RC:$src1), + (_.VT _.RC:$src2), + (i8 imm:$src3), + (i32 FROUND_NO_EXC))>, EVEX_B; +} +//handle scalar instruction reg_vec1 = op(reg_vec2,reg_vec3,imm),{sae} +multiclass avx512_fp_sae_scalar_imm<bits<8> opc, string OpcodeStr, + SDNode OpNode, X86VectorVTInfo _> { + defm NAME: avx512_fp_sae_packed_imm<opc, OpcodeStr, OpNode, _>; +} + +multiclass avx512_common_fp_sae_packed_imm<string OpcodeStr, + AVX512VLVectorVTInfo _, bits<8> opc, SDNode OpNode, Predicate prd>{ + let Predicates = [prd] in { + defm Z : avx512_fp_packed_imm<opc, OpcodeStr, OpNode, _.info512>, + avx512_fp_sae_packed_imm<opc, OpcodeStr, OpNode, _.info512>, + EVEX_V512; + + } + let Predicates = [prd, HasVLX] in { + defm Z128 : avx512_fp_packed_imm<opc, OpcodeStr, OpNode, _.info128>, + EVEX_V128; + defm Z256 : avx512_fp_packed_imm<opc, OpcodeStr, OpNode, _.info256>, + EVEX_V256; + } +} + +multiclass avx512_common_3Op_imm8<string OpcodeStr, AVX512VLVectorVTInfo _, + bits<8> opc, SDNode OpNode>{ + let Predicates = [HasAVX512] in { + defm Z : avx512_3Op_imm8<opc, OpcodeStr, OpNode, _.info512>, EVEX_V512; + } + let Predicates = [HasAVX512, HasVLX] in { + defm Z128 : avx512_3Op_imm8<opc, OpcodeStr, OpNode, _.info128>, EVEX_V128; + defm Z256 : avx512_3Op_imm8<opc, OpcodeStr, OpNode, _.info256>, EVEX_V256; + } +} + +multiclass avx512_common_fp_sae_scalar_imm<string OpcodeStr, + X86VectorVTInfo _, bits<8> opc, SDNode OpNode, Predicate prd>{ + let Predicates = [prd] in { + defm Z128 : avx512_fp_scalar_imm<opc, OpcodeStr, OpNode, _>, + avx512_fp_sae_scalar_imm<opc, OpcodeStr, OpNode, _>; + } +} + +defm VFIXUPIMMPD : avx512_common_fp_sae_packed_imm<"vfixupimmpd", + avx512vl_f64_info, 0x54, X86VFixupimm, HasAVX512>, + AVX512AIi8Base, EVEX_4V, EVEX_CD8<64, CD8VF>, VEX_W; +defm VFIXUPIMMPS : avx512_common_fp_sae_packed_imm<"vfixupimmps", + avx512vl_f32_info, 0x54, X86VFixupimm, HasAVX512>, + AVX512AIi8Base, EVEX_4V, EVEX_CD8<32, CD8VF>; + +defm VFIXUPIMMSD: avx512_common_fp_sae_scalar_imm<"vfixupimmsd", f64x_info, + 0x55, X86VFixupimm, HasAVX512>, + AVX512AIi8Base, VEX_LIG, EVEX_4V, EVEX_CD8<64, CD8VT1>, VEX_W; +defm VFIXUPIMMSS: avx512_common_fp_sae_scalar_imm<"vfixupimmss", f32x_info, + 0x55, X86VFixupimm, HasAVX512>, + AVX512AIi8Base, VEX_LIG, EVEX_4V, EVEX_CD8<32, CD8VT1>; + +defm VRANGEPD : avx512_common_fp_sae_packed_imm<"vrangepd", avx512vl_f64_info, + 0x50, X86VRange, HasDQI>, + AVX512AIi8Base, EVEX_4V, EVEX_CD8<64, CD8VF>, VEX_W; +defm VRANGEPS : avx512_common_fp_sae_packed_imm<"vrangeps", avx512vl_f32_info, + 0x50, X86VRange, HasDQI>, + AVX512AIi8Base, EVEX_4V, EVEX_CD8<32, CD8VF>; + +defm VRANGESD: avx512_common_fp_sae_scalar_imm<"vrangesd", f64x_info, + 0x51, X86VRange, HasDQI>, + AVX512AIi8Base, VEX_LIG, EVEX_4V, EVEX_CD8<64, CD8VT1>, VEX_W; +defm VRANGESS: avx512_common_fp_sae_scalar_imm<"vrangess", f32x_info, + 0x51, X86VRange, HasDQI>, + AVX512AIi8Base, VEX_LIG, EVEX_4V, EVEX_CD8<32, CD8VT1>; + + +multiclass avx512_shuff_packed_128<string OpcodeStr, AVX512VLVectorVTInfo _, + bits<8> opc, SDNode OpNode = X86Shuf128>{ + let Predicates = [HasAVX512] in { + defm Z : avx512_3Op_imm8<opc, OpcodeStr, OpNode, _.info512>, EVEX_V512; + + } + let Predicates = [HasAVX512, HasVLX] in { + defm Z256 : avx512_3Op_imm8<opc, OpcodeStr, OpNode, _.info256>, EVEX_V256; + } +} + +defm VSHUFF32X4 : avx512_shuff_packed_128<"vshuff32x4",avx512vl_f32_info, 0x23>, + AVX512AIi8Base, EVEX_4V, EVEX_CD8<32, CD8VF>; +defm VSHUFF64X2 : avx512_shuff_packed_128<"vshuff64x2",avx512vl_f64_info, 0x23>, + AVX512AIi8Base, EVEX_4V, EVEX_CD8<64, CD8VF>, VEX_W; +defm VSHUFI32X4 : avx512_shuff_packed_128<"vshufi32x4",avx512vl_i32_info, 0x43>, + AVX512AIi8Base, EVEX_4V, EVEX_CD8<32, CD8VF>; +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>{ + 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>, + EVEX_CD8<32, CD8VF>; +defm VALIGNQ: avx512_valign<"valignq", avx512vl_i64_info, avx512vl_f64_info>, + EVEX_CD8<64, CD8VF>, VEX_W; |
