X-Git-Url: http://plrg.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTarget%2FX86%2FX86InstrAVX512.td;h=85a53e1179330390dacafd504478d47ba418555f;hb=2fe1f6e4f3f58218bbcc3568be435745c980c345;hp=ccbe1f4191243b22d5b91e3529eeb71088a38695;hpb=fe11ff1d50fc731c9a7ad6de02dac056fcc17e37;p=oota-llvm.git diff --git a/lib/Target/X86/X86InstrAVX512.td b/lib/Target/X86/X86InstrAVX512.td index ccbe1f41912..85a53e11793 100644 --- a/lib/Target/X86/X86InstrAVX512.td +++ b/lib/Target/X86/X86InstrAVX512.td @@ -1109,62 +1109,6 @@ defm VPBROADCASTMB2Q : avx512_mask_broadcast<0x2A, "vpbroadcastmb2q", } //===----------------------------------------------------------------------===// -// AVX-512 - VPERM -// -// -- immediate form -- -multiclass avx512_perm_imm opc, string OpcodeStr, SDNode OpNode, - X86VectorVTInfo _> { - let ExeDomain = _.ExeDomain in { - def ri : AVX512AIi8, - EVEX; - def mi : AVX512AIi8, - EVEX, EVEX_CD8<_.EltSize, CD8VF>; -} -} - -multiclass avx512_permil OpcImm, bits<8> OpcVar, X86VectorVTInfo _, - X86VectorVTInfo Ctrl> : - avx512_perm_imm { - let ExeDomain = _.ExeDomain in { - def rr : AVX5128I, - EVEX_4V; - def rm : AVX5128I, - EVEX_4V; - } -} -defm VPERMILPSZ : avx512_permil<0x04, 0x0C, v16f32_info, v16i32_info>, - EVEX_V512; -defm VPERMILPDZ : avx512_permil<0x05, 0x0D, v8f64_info, v8i64_info>, - EVEX_V512, VEX_W; - -def : Pat<(v16i32 (X86VPermilpi VR512:$src1, (i8 imm:$imm))), - (VPERMILPSZri VR512:$src1, imm:$imm)>; -def : Pat<(v8i64 (X86VPermilpi VR512:$src1, (i8 imm:$imm))), - (VPERMILPDZri VR512:$src1, imm:$imm)>; - // -- VPERM2I - 3 source operands form -- multiclass avx512_perm_3src opc, string OpcodeStr, SDNode OpNode, X86VectorVTInfo _> { @@ -1363,41 +1307,85 @@ def : Pat<(v8i32 (vselect (v8i1 VK8WM:$mask), (v8i32 VR256X:$src1), //===----------------------------------------------------------------------===// // avx512_cmp_scalar - AVX512 CMPSS and CMPSD -multiclass avx512_cmp_scalar { - def rr : AVX512Ii8<0xC2, MRMSrcReg, - (outs VK1:$dst), (ins RC:$src1, RC:$src2, AVXCC:$cc), - !strconcat("vcmp${cc}", Suffix, + +multiclass avx512_cmp_scalar{ + + defm rr_Int : AVX512_maskable_cmp<0xC2, MRMSrcReg, _, + (outs _.KRC:$dst), + (ins _.RC:$src1, _.RC:$src2, AVXCC:$cc), + "vcmp${cc}"#_.Suffix, + "$src2, $src1", "$src1, $src2", + (OpNode (_.VT _.RC:$src1), + (_.VT _.RC:$src2), + imm:$cc)>, EVEX_4V; + let mayLoad = 1 in + defm rm_Int : AVX512_maskable_cmp<0xC2, MRMSrcMem, _, + (outs _.KRC:$dst), + (ins _.RC:$src1, _.MemOp:$src2, AVXCC:$cc), + "vcmp${cc}"#_.Suffix, + "$src2, $src1", "$src1, $src2", + (OpNode (_.VT _.RC:$src1), + (_.VT (scalar_to_vector (_.ScalarLdFrag addr:$src2))), + imm:$cc)>, EVEX_4V, EVEX_CD8<_.EltSize, CD8VT1>; + + defm rrb_Int : AVX512_maskable_cmp<0xC2, MRMSrcReg, _, + (outs _.KRC:$dst), + (ins _.RC:$src1, _.RC:$src2, AVXCC:$cc), + "vcmp${cc}"#_.Suffix, + "{sae}, $src2, $src1", "$src1, $src2,{sae}", + (OpNodeRnd (_.VT _.RC:$src1), + (_.VT _.RC:$src2), + imm:$cc, + (i32 FROUND_NO_EXC))>, EVEX_4V, EVEX_B; + // Accept explicit immediate argument form instead of comparison code. + let isAsmParserOnly = 1, hasSideEffects = 0 in { + defm rri_alt : AVX512_maskable_cmp_alt<0xC2, MRMSrcReg, _, + (outs VK1:$dst), + (ins _.RC:$src1, _.RC:$src2, u8imm:$cc), + "vcmp"#_.Suffix, + "$cc, $src2, $src1", "$src1, $src2, $cc">, EVEX_4V; + defm rmi_alt : AVX512_maskable_cmp_alt<0xC2, MRMSrcMem, _, + (outs _.KRC:$dst), + (ins _.RC:$src1, _.MemOp:$src2, u8imm:$cc), + "vcmp"#_.Suffix, + "$cc, $src2, $src1", "$src1, $src2, $cc">, + EVEX_4V, EVEX_CD8<_.EltSize, CD8VT1>; + + defm rrb_alt : AVX512_maskable_cmp_alt<0xC2, MRMSrcReg, _, + (outs _.KRC:$dst), + (ins _.RC:$src1, _.RC:$src2, u8imm:$cc), + "vcmp"#_.Suffix, + "$cc,{sae}, $src2, $src1","$src1, $src2,{sae}, $cc">, + EVEX_4V, EVEX_B; + }// let isAsmParserOnly = 1, hasSideEffects = 0 + + let isCodeGenOnly = 1 in { + def rr : AVX512Ii8<0xC2, MRMSrcReg, + (outs _.KRC:$dst), (ins _.FRC:$src1, _.FRC:$src2, AVXCC:$cc), + !strconcat("vcmp${cc}", _.Suffix, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"), - [(set VK1:$dst, (OpNode (VT RC:$src1), RC:$src2, imm:$cc))], + [(set _.KRC:$dst, (OpNode _.FRC:$src1, + _.FRC:$src2, + imm:$cc))], IIC_SSE_ALU_F32S_RR>, EVEX_4V; - def rm : AVX512Ii8<0xC2, MRMSrcMem, - (outs VK1:$dst), (ins RC:$src1, x86memop:$src2, AVXCC:$cc), - !strconcat("vcmp${cc}", Suffix, - "\t{$src2, $src1, $dst|$dst, $src1, $src2}"), - [(set VK1:$dst, (OpNode (VT RC:$src1), - (ld_frag addr:$src2), imm:$cc))], IIC_SSE_ALU_F32P_RM>, EVEX_4V; - let isAsmParserOnly = 1, hasSideEffects = 0 in { - def rri_alt : AVX512Ii8<0xC2, MRMSrcReg, - (outs VK1:$dst), (ins RC:$src1, RC:$src2, u8imm:$cc), - !strconcat("vcmp", Suffix, - "\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}"), - [], IIC_SSE_ALU_F32S_RR>, EVEX_4V; let mayLoad = 1 in - def rmi_alt : AVX512Ii8<0xC2, MRMSrcMem, - (outs VK1:$dst), (ins RC:$src1, x86memop:$src2, u8imm:$cc), - !strconcat("vcmp", Suffix, - "\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}"), - [], IIC_SSE_ALU_F32P_RM>, EVEX_4V; + def rm : AVX512Ii8<0xC2, MRMSrcMem, + (outs _.KRC:$dst), + (ins _.FRC:$src1, _.ScalarMemOp:$src2, AVXCC:$cc), + !strconcat("vcmp${cc}", _.Suffix, + "\t{$src2, $src1, $dst|$dst, $src1, $src2}"), + [(set _.KRC:$dst, (OpNode _.FRC:$src1, + (_.ScalarLdFrag addr:$src2), + imm:$cc))], + IIC_SSE_ALU_F32P_RM>, EVEX_4V, EVEX_CD8<_.EltSize, CD8VT1>; } } let Predicates = [HasAVX512] in { -defm VCMPSSZ : avx512_cmp_scalar, - XS; -defm VCMPSDZ : avx512_cmp_scalar, - XD, VEX_W; + defm VCMPSSZ : avx512_cmp_scalar, + AVX512XSIi8Base; + defm VCMPSDZ : avx512_cmp_scalar, + AVX512XDIi8Base, VEX_W; } multiclass avx512_icmp_packed opc, string OpcodeStr, SDNode OpNode, @@ -1854,7 +1842,6 @@ multiclass avx512_vector_fpclass opc, string OpcodeStr, SDNode OpNode, } } - multiclass avx512_vector_fpclass_all opc, SDNode OpNode, Predicate prd, string broadcast>{ @@ -2023,10 +2010,15 @@ let Predicates = [HasAVX512] in { (AND32ri (KMOVWrk (COPY_TO_REGCLASS VK1:$src, VK16)), (i32 1))>; def : Pat<(i32 (anyext VK1:$src)), (KMOVWrk (COPY_TO_REGCLASS VK1:$src, VK16))>; + def : Pat<(i8 (zext VK1:$src)), (EXTRACT_SUBREG (AND32ri (KMOVWrk (COPY_TO_REGCLASS VK1:$src, VK16)), (i32 1)), sub_8bit)>; + def : Pat<(i8 (anyext VK1:$src)), + (EXTRACT_SUBREG + (KMOVWrk (COPY_TO_REGCLASS VK1:$src, VK16)), sub_8bit)>; + def : Pat<(i64 (zext VK1:$src)), (AND64ri8 (SUBREG_TO_REG (i64 0), (KMOVWrk (COPY_TO_REGCLASS VK1:$src, VK16)), sub_32bit), (i64 1))>; @@ -4081,7 +4073,73 @@ defm VPERMQ : avx512_vpermi_dq_sizes<0x00, MRMSrcReg, MRMSrcMem, "vpermq", defm VPERMPD : avx512_vpermi_dq_sizes<0x01, MRMSrcReg, MRMSrcMem, "vpermpd", X86VPermi, avx512vl_f64_info>, EVEX, AVX512AIi8Base, EVEX_CD8<64, CD8VF>, VEX_W; +//===----------------------------------------------------------------------===// +// AVX-512 - VPERMIL +//===----------------------------------------------------------------------===// +multiclass avx512_permil_vec OpcVar, string OpcodeStr, SDNode OpNode, + X86VectorVTInfo _, X86VectorVTInfo Ctrl> { + defm rr: AVX512_maskable, + T8PD, EVEX_4V; + let mayLoad = 1 in { + defm rm: AVX512_maskable, + T8PD, EVEX_4V, EVEX_CD8<_.EltSize, CD8VF>; + defm rmb: AVX512_maskable, + T8PD, EVEX_4V, EVEX_B, EVEX_CD8<_.EltSize, CD8VF>; + }//let mayLoad = 1 +} + +multiclass avx512_permil_vec_common OpcVar, + AVX512VLVectorVTInfo _, AVX512VLVectorVTInfo Ctrl>{ + let Predicates = [HasAVX512] in { + defm Z : avx512_permil_vec, EVEX_V512; + } + let Predicates = [HasAVX512, HasVLX] in { + defm Z128 : avx512_permil_vec, EVEX_V128; + defm Z256 : avx512_permil_vec, EVEX_V256; + } +} + +multiclass avx512_permil OpcImm, bits<8> OpcVar, + AVX512VLVectorVTInfo _, AVX512VLVectorVTInfo Ctrl>{ + + defm NAME: avx512_permil_vec_common; + defm NAME: avx512_shift_rmi_sizes, + EVEX, AVX512AIi8Base, EVEX_CD8<_.info128.EltSize, CD8VF>; + + let isCodeGenOnly = 1 in { + // lowering implementation with the alternative types + defm NAME#_I: avx512_permil_vec_common; + defm NAME#_I: avx512_shift_rmi_sizes, + EVEX, AVX512AIi8Base, EVEX_CD8<_.info128.EltSize, CD8VF>; + } +} + +defm VPERMILPS : avx512_permil<"vpermilps", 0x04, 0x0C, avx512vl_f32_info, + avx512vl_i32_info>; +defm VPERMILPD : avx512_permil<"vpermilpd", 0x05, 0x0D, avx512vl_f64_info, + avx512vl_i64_info>, VEX_W; //===----------------------------------------------------------------------===// // AVX-512 - VPSHUFD, VPSHUFLW, VPSHUFHW //===----------------------------------------------------------------------===// @@ -4090,9 +4148,9 @@ 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; + X86PShufhw>, EVEX, AVX512XSIi8Base; defm VPSHUFL : avx512_shift_rmi_w<0x70, MRMSrcReg, MRMSrcMem, "vpshuflw", - X86PShuflw>, EVEX, AVX512XDIi8Base, VEX_W; + X86PShuflw>, EVEX, AVX512XDIi8Base; multiclass avx512_pshufb_sizes opc, string OpcodeStr, SDNode OpNode> { let Predicates = [HasBWI] in @@ -5311,50 +5369,31 @@ let Defs = [EFLAGS], Predicates = [HasAVX512] in { } /// avx512_fp14_s rcp14ss, rcp14sd, rsqrt14ss, rsqrt14sd -multiclass avx512_fp14_s opc, string OpcodeStr, RegisterClass RC, - X86MemOperand x86memop> { - let hasSideEffects = 0 in { - def rr : AVX5128I, EVEX_4V; +multiclass avx512_fp14_s opc, string OpcodeStr, SDNode OpNode, + X86VectorVTInfo _> { + let hasSideEffects = 0, AddedComplexity = 20 , Predicates = [HasAVX512] in { + defm rr : AVX512_maskable_scalar, EVEX_4V; let mayLoad = 1 in { - def rm : AVX5128I, EVEX_4V; + defm rm : AVX512_maskable_scalar, EVEX_4V; } } } -defm VRCP14SS : avx512_fp14_s<0x4D, "vrcp14ss", FR32X, f32mem>, - EVEX_CD8<32, CD8VT1>; -defm VRCP14SD : avx512_fp14_s<0x4D, "vrcp14sd", FR64X, f64mem>, - VEX_W, EVEX_CD8<64, CD8VT1>; -defm VRSQRT14SS : avx512_fp14_s<0x4F, "vrsqrt14ss", FR32X, f32mem>, - EVEX_CD8<32, CD8VT1>; -defm VRSQRT14SD : avx512_fp14_s<0x4F, "vrsqrt14sd", FR64X, f64mem>, - VEX_W, EVEX_CD8<64, CD8VT1>; - -def : Pat <(v4f32 (int_x86_avx512_rcp14_ss (v4f32 VR128X:$src1), - (v4f32 VR128X:$src2), (bc_v4f32 (v4i32 immAllZerosV)), (i8 -1))), - (COPY_TO_REGCLASS (VRCP14SSrr (COPY_TO_REGCLASS VR128X:$src1, FR32X), - (COPY_TO_REGCLASS VR128X:$src2, FR32X)), VR128X)>; - -def : Pat <(v2f64 (int_x86_avx512_rcp14_sd (v2f64 VR128X:$src1), - (v2f64 VR128X:$src2), (bc_v2f64 (v4i32 immAllZerosV)), (i8 -1))), - (COPY_TO_REGCLASS (VRCP14SDrr (COPY_TO_REGCLASS VR128X:$src1, FR64X), - (COPY_TO_REGCLASS VR128X:$src2, FR64X)), VR128X)>; - -def : Pat <(v4f32 (int_x86_avx512_rsqrt14_ss (v4f32 VR128X:$src1), - (v4f32 VR128X:$src2), (bc_v4f32 (v4i32 immAllZerosV)), (i8 -1))), - (COPY_TO_REGCLASS (VRSQRT14SSrr (COPY_TO_REGCLASS VR128X:$src1, FR32X), - (COPY_TO_REGCLASS VR128X:$src2, FR32X)), VR128X)>; - -def : Pat <(v2f64 (int_x86_avx512_rsqrt14_sd (v2f64 VR128X:$src1), - (v2f64 VR128X:$src2), (bc_v2f64 (v4i32 immAllZerosV)), (i8 -1))), - (COPY_TO_REGCLASS (VRSQRT14SDrr (COPY_TO_REGCLASS VR128X:$src1, FR64X), - (COPY_TO_REGCLASS VR128X:$src2, FR64X)), VR128X)>; +defm VRCP14SS : avx512_fp14_s<0x4D, "vrcp14ss", X86frcp14s, f32x_info>, + EVEX_CD8<32, CD8VT1>, T8PD; +defm VRCP14SD : avx512_fp14_s<0x4D, "vrcp14sd", X86frcp14s, f64x_info>, + VEX_W, EVEX_CD8<64, CD8VT1>, T8PD; +defm VRSQRT14SS : avx512_fp14_s<0x4F, "vrsqrt14ss", X86frsqrt14s, f32x_info>, + EVEX_CD8<32, CD8VT1>, T8PD; +defm VRSQRT14SD : avx512_fp14_s<0x4F, "vrsqrt14sd", X86frsqrt14s, f64x_info>, + VEX_W, EVEX_CD8<64, CD8VT1>, T8PD; /// avx512_fp14_p rcp14ps, rcp14pd, rsqrt14ps, rsqrt14pd multiclass avx512_fp14_p opc, string OpcodeStr, SDNode OpNode, @@ -5636,15 +5675,14 @@ defm VSQRT : avx512_sqrt_scalar_all<0x51, "vsqrt">, VEX_LIG; let Predicates = [HasAVX512] in { def : Pat<(f32 (X86frsqrt FR32X:$src)), - (VRSQRT14SSrr (f32 (IMPLICIT_DEF)), FR32X:$src)>; + (COPY_TO_REGCLASS (VRSQRT14SSrr (v4f32 (IMPLICIT_DEF)), (COPY_TO_REGCLASS FR32X:$src, VR128X)), VR128X)>; def : Pat<(f32 (X86frsqrt (load addr:$src))), - (VRSQRT14SSrm (f32 (IMPLICIT_DEF)), addr:$src)>, + (COPY_TO_REGCLASS (VRSQRT14SSrm (v4f32 (IMPLICIT_DEF)), addr:$src), VR128X)>, Requires<[OptForSize]>; - def : Pat<(f32 (X86frcp FR32X:$src)), - (VRCP14SSrr (f32 (IMPLICIT_DEF)), FR32X:$src)>; + (COPY_TO_REGCLASS (VRCP14SSrr (v4f32 (IMPLICIT_DEF)), (COPY_TO_REGCLASS FR32X:$src, VR128X)), VR128X )>; def : Pat<(f32 (X86frcp (load addr:$src))), - (VRCP14SSrm (f32 (IMPLICIT_DEF)), addr:$src)>, + (COPY_TO_REGCLASS (VRCP14SSrm (v4f32 (IMPLICIT_DEF)), addr:$src), VR128X)>, Requires<[OptForSize]>; }