X-Git-Url: http://plrg.eecs.uci.edu/git/?p=oota-llvm.git;a=blobdiff_plain;f=lib%2FTarget%2FX86%2FX86InstrAVX512.td;h=1d1abcfb40fbd34dee294d5bbbe72dd38bfbf495;hp=34d517fb72ef0543a2de814148355e2de716253b;hb=6bc8d95153d1f311a02cf48ca0e5735c80a531a3;hpb=0d5d656524e11532c3f8fbd3c0deae9236676bf7 diff --git a/lib/Target/X86/X86InstrAVX512.td b/lib/Target/X86/X86InstrAVX512.td index 34d517fb72e..1d1abcfb40f 100644 --- a/lib/Target/X86/X86InstrAVX512.td +++ b/lib/Target/X86/X86InstrAVX512.td @@ -1,19 +1,266 @@ +// Group template arguments that can be derived from the vector type (EltNum x +// EltVT). These are things like the register class for the writemask, etc. +// The idea is to pass one of these as the template argument rather than the +// individual arguments. +class X86VectorVTInfo { + RegisterClass RC = rc; + int NumElts = numelts; + + // Corresponding mask register class. + RegisterClass KRC = !cast("VK" # NumElts); + + // Corresponding write-mask register class. + RegisterClass KRCWM = !cast("VK" # NumElts # "WM"); + + // The GPR register class that can hold the write mask. Use GR8 for fewer + // than 8 elements. Use shift-right and equal to work around the lack of + // !lt in tablegen. + RegisterClass MRC = + !cast("GR" # + !if (!eq (!srl(NumElts, 3), 0), 8, NumElts)); + + // Suffix used in the instruction mnemonic. + string Suffix = suffix; + + string VTName = "v" # NumElts # EltVT; + + // The vector VT. + ValueType VT = !cast(VTName); + + string EltTypeName = !cast(EltVT); + // Size of the element type in bits, e.g. 32 for v16i32. + string EltSizeName = !subst("i", "", !subst("f", "", EltTypeName)); + int EltSize = EltVT.Size; + + // "i" for integer types and "f" for floating-point types + string TypeVariantName = !subst(EltSizeName, "", EltTypeName); + + // Size of RC in bits, e.g. 512 for VR512. + int Size = VT.Size; + + // The corresponding memory operand, e.g. i512mem for VR512. + X86MemOperand MemOp = !cast(TypeVariantName # Size # "mem"); + X86MemOperand ScalarMemOp = !cast(EltVT # "mem"); + + // Load patterns + // Note: For 128/256-bit integer VT we choose loadv2i64/loadv4i64 + // due to load promotion during legalization + PatFrag LdFrag = !cast("load" # + !if (!eq (TypeVariantName, "i"), + !if (!eq (Size, 128), "v2i64", + !if (!eq (Size, 256), "v4i64", + VTName)), VTName)); + PatFrag ScalarLdFrag = !cast("load" # EltVT); + + // Load patterns used for memory operands. We only have this defined in + // case of i64 element types for sub-512 integer vectors. For now, keep + // MemOpFrag undefined in these cases. + PatFrag MemOpFrag = + !if (!eq (TypeVariantName, "f"), !cast("memop" # VTName), + !if (!eq (EltTypeName, "i64"), !cast("memop" # VTName), + !if (!eq (VTName, "v16i32"), !cast("memop" # VTName), ?))); + + // The corresponding float type, e.g. v16f32 for v16i32 + // Note: For EltSize < 32, FloatVT is illegal and TableGen + // fails to compile, so we choose FloatVT = VT + ValueType FloatVT = !cast( + !if (!eq (!srl(EltSize,5),0), + VTName, + !if (!eq(TypeVariantName, "i"), + "v" # NumElts # "f" # EltSize, + VTName))); + + // The string to specify embedded broadcast in assembly. + string BroadcastStr = "{1to" # NumElts # "}"; + + // 8-bit compressed displacement tuple/subvector format. This is only + // defined for NumElts <= 8. + CD8VForm CD8TupleForm = !if (!eq (!srl(NumElts, 4), 0), + !cast("CD8VT" # NumElts), ?); + + SubRegIndex SubRegIdx = !if (!eq (Size, 128), sub_xmm, + !if (!eq (Size, 256), sub_ymm, ?)); + + Domain ExeDomain = !if (!eq (EltTypeName, "f32"), SSEPackedSingle, + !if (!eq (EltTypeName, "f64"), SSEPackedDouble, + SSEPackedInt)); + + // A vector type of the same width with element type i32. This is used to + // create the canonical constant zero node ImmAllZerosV. + ValueType i32VT = !cast("v" # !srl(Size, 5) # "i32"); + dag ImmAllZerosV = (VT (bitconvert (i32VT immAllZerosV))); +} + +def v64i8_info : X86VectorVTInfo<64, i8, VR512, "b">; +def v32i16_info : X86VectorVTInfo<32, i16, VR512, "w">; +def v16i32_info : X86VectorVTInfo<16, i32, VR512, "d">; +def v8i64_info : X86VectorVTInfo<8, i64, VR512, "q">; +def v16f32_info : X86VectorVTInfo<16, f32, VR512, "ps">; +def v8f64_info : X86VectorVTInfo<8, f64, VR512, "pd">; + +// "x" in v32i8x_info means RC = VR256X +def v32i8x_info : X86VectorVTInfo<32, i8, VR256X, "b">; +def v16i16x_info : X86VectorVTInfo<16, i16, VR256X, "w">; +def v8i32x_info : X86VectorVTInfo<8, i32, VR256X, "d">; +def v4i64x_info : X86VectorVTInfo<4, i64, VR256X, "q">; + +def v16i8x_info : X86VectorVTInfo<16, i8, VR128X, "b">; +def v8i16x_info : X86VectorVTInfo<8, i16, VR128X, "w">; +def v4i32x_info : X86VectorVTInfo<4, i32, VR128X, "d">; +def v2i64x_info : X86VectorVTInfo<2, i64, VR128X, "q">; + +class AVX512VLVectorVTInfo { + X86VectorVTInfo info512 = i512; + X86VectorVTInfo info256 = i256; + X86VectorVTInfo info128 = i128; +} + +def avx512vl_i8_info : AVX512VLVectorVTInfo; +def avx512vl_i16_info : AVX512VLVectorVTInfo; +def avx512vl_i32_info : AVX512VLVectorVTInfo; +def avx512vl_i64_info : AVX512VLVectorVTInfo; + +// This multiclass generates the masking variants from the non-masking +// variant. It only provides the assembly pieces for the masking variants. +// It assumes custom ISel patterns for masking which can be provided as +// template arguments. +multiclass AVX512_maskable_custom O, Format F, + dag Outs, + dag Ins, dag MaskingIns, dag ZeroMaskingIns, + string OpcodeStr, + string AttSrcAsm, string IntelSrcAsm, + list Pattern, + list MaskingPattern, + list ZeroMaskingPattern, + string MaskingConstraint = "", + InstrItinClass itin = NoItinerary, + bit IsCommutable = 0> { + let isCommutable = IsCommutable in + def NAME: AVX512; + + // Prefer over VMOV*rrk Pat<> + let AddedComplexity = 20 in + def NAME#k: AVX512, + EVEX_K { + // In case of the 3src subclass this is overridden with a let. + string Constraints = MaskingConstraint; + } + let AddedComplexity = 30 in // Prefer over VMOV*rrkz Pat<> + def NAME#kz: AVX512, + EVEX_KZ; +} + + +// Common base class of AVX512_maskable and AVX512_maskable_3src. +multiclass AVX512_maskable_common O, Format F, X86VectorVTInfo _, + dag Outs, + dag Ins, dag MaskingIns, dag ZeroMaskingIns, + string OpcodeStr, + string AttSrcAsm, string IntelSrcAsm, + dag RHS, dag MaskingRHS, + string MaskingConstraint = "", + InstrItinClass itin = NoItinerary, + bit IsCommutable = 0> : + AVX512_maskable_custom; + +// This multiclass generates the unconditional/non-masking, the masking and +// the zero-masking variant of the instruction. In the masking case, the +// perserved vector elements come from a new dummy input operand tied to $dst. +multiclass AVX512_maskable O, Format F, X86VectorVTInfo _, + dag Outs, dag Ins, string OpcodeStr, + string AttSrcAsm, string IntelSrcAsm, + dag RHS, InstrItinClass itin = NoItinerary, + bit IsCommutable = 0> : + AVX512_maskable_common; + +// Similar to AVX512_maskable but in this case one of the source operands +// ($src1) is already tied to $dst so we just use that for the preserved +// vector elements. NOTE that the NonTiedIns (the ins dag) should exclude +// $src1. +multiclass AVX512_maskable_3src O, Format F, X86VectorVTInfo _, + dag Outs, dag NonTiedIns, string OpcodeStr, + string AttSrcAsm, string IntelSrcAsm, + dag RHS> : + AVX512_maskable_common; + + +multiclass AVX512_maskable_in_asm O, Format F, X86VectorVTInfo _, + dag Outs, dag Ins, + string OpcodeStr, + string AttSrcAsm, string IntelSrcAsm, + list Pattern> : + AVX512_maskable_custom; + // Bitcasts between 512-bit vector types. Return the original type since // no instruction is needed for the conversion let Predicates = [HasAVX512] in { - def : Pat<(v8f64 (bitconvert (v16f32 VR512:$src))), (v8f64 VR512:$src)>; - def : Pat<(v8f64 (bitconvert (v16i32 VR512:$src))), (v8f64 VR512:$src)>; def : Pat<(v8f64 (bitconvert (v8i64 VR512:$src))), (v8f64 VR512:$src)>; - def : Pat<(v16f32 (bitconvert (v16i32 VR512:$src))), (v16f32 VR512:$src)>; + def : Pat<(v8f64 (bitconvert (v16i32 VR512:$src))), (v8f64 VR512:$src)>; + def : Pat<(v8f64 (bitconvert (v32i16 VR512:$src))), (v8f64 VR512:$src)>; + def : Pat<(v8f64 (bitconvert (v64i8 VR512:$src))), (v8f64 VR512:$src)>; + def : Pat<(v8f64 (bitconvert (v16f32 VR512:$src))), (v8f64 VR512:$src)>; def : Pat<(v16f32 (bitconvert (v8i64 VR512:$src))), (v16f32 VR512:$src)>; + def : Pat<(v16f32 (bitconvert (v16i32 VR512:$src))), (v16f32 VR512:$src)>; + def : Pat<(v16f32 (bitconvert (v32i16 VR512:$src))), (v16f32 VR512:$src)>; + def : Pat<(v16f32 (bitconvert (v64i8 VR512:$src))), (v16f32 VR512:$src)>; def : Pat<(v16f32 (bitconvert (v8f64 VR512:$src))), (v16f32 VR512:$src)>; - def : Pat<(v8i64 (bitconvert (v16f32 VR512:$src))), (v8i64 VR512:$src)>; def : Pat<(v8i64 (bitconvert (v16i32 VR512:$src))), (v8i64 VR512:$src)>; + def : Pat<(v8i64 (bitconvert (v32i16 VR512:$src))), (v8i64 VR512:$src)>; + def : Pat<(v8i64 (bitconvert (v64i8 VR512:$src))), (v8i64 VR512:$src)>; def : Pat<(v8i64 (bitconvert (v8f64 VR512:$src))), (v8i64 VR512:$src)>; + def : Pat<(v8i64 (bitconvert (v16f32 VR512:$src))), (v8i64 VR512:$src)>; + def : Pat<(v16i32 (bitconvert (v8i64 VR512:$src))), (v16i32 VR512:$src)>; def : Pat<(v16i32 (bitconvert (v16f32 VR512:$src))), (v16i32 VR512:$src)>; - def : Pat<(v16i32 (bitconvert (v8i64 VR512:$src))), (v16i32 VR512:$src)>; + def : Pat<(v16i32 (bitconvert (v32i16 VR512:$src))), (v16i32 VR512:$src)>; + def : Pat<(v16i32 (bitconvert (v64i8 VR512:$src))), (v16i32 VR512:$src)>; def : Pat<(v16i32 (bitconvert (v8f64 VR512:$src))), (v16i32 VR512:$src)>; - def : Pat<(v8f64 (bitconvert (v8i64 VR512:$src))), (v8f64 VR512:$src)>; + def : Pat<(v32i16 (bitconvert (v8i64 VR512:$src))), (v32i16 VR512:$src)>; + def : Pat<(v32i16 (bitconvert (v16i32 VR512:$src))), (v32i16 VR512:$src)>; + def : Pat<(v32i16 (bitconvert (v64i8 VR512:$src))), (v32i16 VR512:$src)>; + def : Pat<(v32i16 (bitconvert (v8f64 VR512:$src))), (v32i16 VR512:$src)>; + def : Pat<(v32i16 (bitconvert (v16f32 VR512:$src))), (v32i16 VR512:$src)>; + def : Pat<(v32i16 (bitconvert (v16f32 VR512:$src))), (v32i16 VR512:$src)>; + def : Pat<(v64i8 (bitconvert (v8i64 VR512:$src))), (v64i8 VR512:$src)>; + def : Pat<(v64i8 (bitconvert (v16i32 VR512:$src))), (v64i8 VR512:$src)>; + def : Pat<(v64i8 (bitconvert (v32i16 VR512:$src))), (v64i8 VR512:$src)>; + def : Pat<(v64i8 (bitconvert (v8f64 VR512:$src))), (v64i8 VR512:$src)>; + def : Pat<(v64i8 (bitconvert (v16f32 VR512:$src))), (v64i8 VR512:$src)>; def : Pat<(v2i64 (bitconvert (v4i32 VR128X:$src))), (v2i64 VR128X:$src)>; def : Pat<(v2i64 (bitconvert (v8i16 VR128X:$src))), (v2i64 VR128X:$src)>; @@ -99,228 +346,184 @@ def : Pat<(v8f64 immAllZerosV), (AVX512_512_SET0)>; //===----------------------------------------------------------------------===// // AVX-512 - VECTOR INSERT // -// -- 32x8 form -- -let hasSideEffects = 0, ExeDomain = SSEPackedSingle in { -def VINSERTF32x4rr : AVX512AIi8<0x18, MRMSrcReg, (outs VR512:$dst), - (ins VR512:$src1, VR128X:$src2, i8imm:$src3), - "vinsertf32x4\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}", - []>, EVEX_4V, EVEX_V512; -let mayLoad = 1 in -def VINSERTF32x4rm : AVX512AIi8<0x18, MRMSrcMem, (outs VR512:$dst), - (ins VR512:$src1, f128mem:$src2, i8imm:$src3), - "vinsertf32x4\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}", - []>, EVEX_4V, EVEX_V512, EVEX_CD8<32, CD8VT4>; -} - -// -- 64x4 fp form -- -let hasSideEffects = 0, ExeDomain = SSEPackedDouble in { -def VINSERTF64x4rr : AVX512AIi8<0x1a, MRMSrcReg, (outs VR512:$dst), - (ins VR512:$src1, VR256X:$src2, i8imm:$src3), - "vinsertf64x4\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}", - []>, EVEX_4V, EVEX_V512, VEX_W; -let mayLoad = 1 in -def VINSERTF64x4rm : AVX512AIi8<0x1a, MRMSrcMem, (outs VR512:$dst), - (ins VR512:$src1, i256mem:$src2, i8imm:$src3), - "vinsertf64x4\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}", - []>, EVEX_4V, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VT4>; + +multiclass vinsert_for_size_no_alt { + let hasSideEffects = 0, ExeDomain = To.ExeDomain in { + def rr : AVX512AIi8, + EVEX_4V, EVEX_V512; + + let mayLoad = 1 in + def rm : AVX512AIi8, + EVEX_4V, EVEX_V512, EVEX_CD8; + } } -// -- 32x4 integer form -- -let hasSideEffects = 0 in { -def VINSERTI32x4rr : AVX512AIi8<0x38, MRMSrcReg, (outs VR512:$dst), - (ins VR512:$src1, VR128X:$src2, i8imm:$src3), - "vinserti32x4\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}", - []>, EVEX_4V, EVEX_V512; -let mayLoad = 1 in -def VINSERTI32x4rm : AVX512AIi8<0x38, MRMSrcMem, (outs VR512:$dst), - (ins VR512:$src1, i128mem:$src2, i8imm:$src3), - "vinserti32x4\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}", - []>, EVEX_4V, EVEX_V512, EVEX_CD8<32, CD8VT4>; +multiclass vinsert_for_size : + vinsert_for_size_no_alt { + // Codegen pattern with the alternative types, e.g. v2i64 -> v8i64 for + // vinserti32x4. Only add this if 64x2 and friends are not supported + // natively via AVX512DQ. + let Predicates = [NoDQI] in + def : Pat<(vinsert_insert:$ins + (AltTo.VT VR512:$src1), (AltFrom.VT From.RC:$src2), (iPTR imm)), + (AltTo.VT (!cast(NAME # From.EltSize # "x4rr") + VR512:$src1, From.RC:$src2, + (INSERT_get_vinsert_imm VR512:$ins)))>; } -let hasSideEffects = 0 in { -// -- 64x4 form -- -def VINSERTI64x4rr : AVX512AIi8<0x3a, MRMSrcReg, (outs VR512:$dst), - (ins VR512:$src1, VR256X:$src2, i8imm:$src3), - "vinserti64x4\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}", - []>, EVEX_4V, EVEX_V512, VEX_W; -let mayLoad = 1 in -def VINSERTI64x4rm : AVX512AIi8<0x3a, MRMSrcMem, (outs VR512:$dst), - (ins VR512:$src1, i256mem:$src2, i8imm:$src3), - "vinserti64x4\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}", - []>, EVEX_4V, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VT4>; -} - -def : Pat<(vinsert128_insert:$ins (v16f32 VR512:$src1), (v4f32 VR128X:$src2), - (iPTR imm)), (VINSERTF32x4rr VR512:$src1, VR128X:$src2, - (INSERT_get_vinsert128_imm VR512:$ins))>; -def : Pat<(vinsert128_insert:$ins (v8f64 VR512:$src1), (v2f64 VR128X:$src2), - (iPTR imm)), (VINSERTF32x4rr VR512:$src1, VR128X:$src2, - (INSERT_get_vinsert128_imm VR512:$ins))>; -def : Pat<(vinsert128_insert:$ins (v8i64 VR512:$src1), (v2i64 VR128X:$src2), - (iPTR imm)), (VINSERTI32x4rr VR512:$src1, VR128X:$src2, - (INSERT_get_vinsert128_imm VR512:$ins))>; -def : Pat<(vinsert128_insert:$ins (v16i32 VR512:$src1), (v4i32 VR128X:$src2), - (iPTR imm)), (VINSERTI32x4rr VR512:$src1, VR128X:$src2, - (INSERT_get_vinsert128_imm VR512:$ins))>; - -def : Pat<(vinsert128_insert:$ins (v16f32 VR512:$src1), (loadv4f32 addr:$src2), - (iPTR imm)), (VINSERTF32x4rm VR512:$src1, addr:$src2, - (INSERT_get_vinsert128_imm VR512:$ins))>; -def : Pat<(vinsert128_insert:$ins (v16i32 VR512:$src1), - (bc_v4i32 (loadv2i64 addr:$src2)), - (iPTR imm)), (VINSERTI32x4rm VR512:$src1, addr:$src2, - (INSERT_get_vinsert128_imm VR512:$ins))>; -def : Pat<(vinsert128_insert:$ins (v8f64 VR512:$src1), (loadv2f64 addr:$src2), - (iPTR imm)), (VINSERTF32x4rm VR512:$src1, addr:$src2, - (INSERT_get_vinsert128_imm VR512:$ins))>; -def : Pat<(vinsert128_insert:$ins (v8i64 VR512:$src1), (loadv2i64 addr:$src2), - (iPTR imm)), (VINSERTI32x4rm VR512:$src1, addr:$src2, - (INSERT_get_vinsert128_imm VR512:$ins))>; - -def : Pat<(vinsert256_insert:$ins (v16f32 VR512:$src1), (v8f32 VR256X:$src2), - (iPTR imm)), (VINSERTF64x4rr VR512:$src1, VR256X:$src2, - (INSERT_get_vinsert256_imm VR512:$ins))>; -def : Pat<(vinsert256_insert:$ins (v8f64 VR512:$src1), (v4f64 VR256X:$src2), - (iPTR imm)), (VINSERTF64x4rr VR512:$src1, VR256X:$src2, - (INSERT_get_vinsert256_imm VR512:$ins))>; -def : Pat<(vinsert128_insert:$ins (v8i64 VR512:$src1), (v4i64 VR256X:$src2), - (iPTR imm)), (VINSERTI64x4rr VR512:$src1, VR256X:$src2, - (INSERT_get_vinsert256_imm VR512:$ins))>; -def : Pat<(vinsert128_insert:$ins (v16i32 VR512:$src1), (v8i32 VR256X:$src2), - (iPTR imm)), (VINSERTI64x4rr VR512:$src1, VR256X:$src2, - (INSERT_get_vinsert256_imm VR512:$ins))>; - -def : Pat<(vinsert256_insert:$ins (v16f32 VR512:$src1), (loadv8f32 addr:$src2), - (iPTR imm)), (VINSERTF64x4rm VR512:$src1, addr:$src2, - (INSERT_get_vinsert256_imm VR512:$ins))>; -def : Pat<(vinsert256_insert:$ins (v8f64 VR512:$src1), (loadv4f64 addr:$src2), - (iPTR imm)), (VINSERTF64x4rm VR512:$src1, addr:$src2, - (INSERT_get_vinsert256_imm VR512:$ins))>; -def : Pat<(vinsert256_insert:$ins (v8i64 VR512:$src1), (loadv4i64 addr:$src2), - (iPTR imm)), (VINSERTI64x4rm VR512:$src1, addr:$src2, - (INSERT_get_vinsert256_imm VR512:$ins))>; -def : Pat<(vinsert256_insert:$ins (v16i32 VR512:$src1), - (bc_v8i32 (loadv4i64 addr:$src2)), - (iPTR imm)), (VINSERTI64x4rm VR512:$src1, addr:$src2, - (INSERT_get_vinsert256_imm VR512:$ins))>; +multiclass vinsert_for_type { + defm NAME # "32x4" : vinsert_for_size, + X86VectorVTInfo<16, EltVT32, VR512>, + X86VectorVTInfo< 2, EltVT64, VR128X>, + X86VectorVTInfo< 8, EltVT64, VR512>, + vinsert128_insert, + INSERT_get_vinsert128_imm>; + let Predicates = [HasDQI] in + defm NAME # "64x2" : vinsert_for_size_no_alt, + X86VectorVTInfo< 8, EltVT64, VR512>, + vinsert128_insert, + INSERT_get_vinsert128_imm>, VEX_W; + defm NAME # "64x4" : vinsert_for_size, + X86VectorVTInfo< 8, EltVT64, VR512>, + X86VectorVTInfo< 8, EltVT32, VR256>, + X86VectorVTInfo<16, EltVT32, VR512>, + vinsert256_insert, + INSERT_get_vinsert256_imm>, VEX_W; + let Predicates = [HasDQI] in + defm NAME # "32x8" : vinsert_for_size_no_alt, + X86VectorVTInfo<16, EltVT32, VR512>, + vinsert256_insert, + INSERT_get_vinsert256_imm>; +} + +defm VINSERTF : vinsert_for_type; +defm VINSERTI : vinsert_for_type; // vinsertps - insert f32 to XMM def VINSERTPSzrr : AVX512AIi8<0x21, MRMSrcReg, (outs VR128X:$dst), - (ins VR128X:$src1, VR128X:$src2, u32u8imm:$src3), + (ins VR128X:$src1, VR128X:$src2, i8imm:$src3), "vinsertps\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}", - [(set VR128X:$dst, (X86insrtps VR128X:$src1, VR128X:$src2, imm:$src3))]>, + [(set VR128X:$dst, (X86insertps VR128X:$src1, VR128X:$src2, imm:$src3))]>, EVEX_4V; def VINSERTPSzrm: AVX512AIi8<0x21, MRMSrcMem, (outs VR128X:$dst), - (ins VR128X:$src1, f32mem:$src2, u32u8imm:$src3), + (ins VR128X:$src1, f32mem:$src2, i8imm:$src3), "vinsertps\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}", - [(set VR128X:$dst, (X86insrtps VR128X:$src1, + [(set VR128X:$dst, (X86insertps VR128X:$src1, (v4f32 (scalar_to_vector (loadf32 addr:$src2))), imm:$src3))]>, EVEX_4V, EVEX_CD8<32, CD8VT1>; //===----------------------------------------------------------------------===// // AVX-512 VECTOR EXTRACT //--- -let hasSideEffects = 0, ExeDomain = SSEPackedSingle in { -// -- 32x4 form -- -def VEXTRACTF32x4rr : AVX512AIi8<0x19, MRMDestReg, (outs VR128X:$dst), - (ins VR512:$src1, i8imm:$src2), - "vextractf32x4\t{$src2, $src1, $dst|$dst, $src1, $src2}", - []>, EVEX, EVEX_V512; -def VEXTRACTF32x4mr : AVX512AIi8<0x19, MRMDestMem, (outs), - (ins f128mem:$dst, VR512:$src1, i8imm:$src2), - "vextractf32x4\t{$src2, $src1, $dst|$dst, $src1, $src2}", - []>, EVEX, EVEX_V512, EVEX_CD8<32, CD8VT4>; - -// -- 64x4 form -- -def VEXTRACTF64x4rr : AVX512AIi8<0x1b, MRMDestReg, (outs VR256X:$dst), - (ins VR512:$src1, i8imm:$src2), - "vextractf64x4\t{$src2, $src1, $dst|$dst, $src1, $src2}", - []>, EVEX, EVEX_V512, VEX_W; -let mayStore = 1 in -def VEXTRACTF64x4mr : AVX512AIi8<0x1b, MRMDestMem, (outs), - (ins f256mem:$dst, VR512:$src1, i8imm:$src2), - "vextractf64x4\t{$src2, $src1, $dst|$dst, $src1, $src2}", - []>, EVEX, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VT4>; + +multiclass vextract_for_size { + let hasSideEffects = 0, ExeDomain = To.ExeDomain in { + defm rr : AVX512_maskable_in_asm, + AVX512AIi8Base, EVEX, EVEX_V512; + let mayStore = 1 in + def rm : AVX512AIi8, EVEX, EVEX_V512, EVEX_CD8; + } + + // Codegen pattern with the alternative types, e.g. v8i64 -> v2i64 for + // vextracti32x4 + def : Pat<(vextract_extract:$ext (AltFrom.VT VR512:$src1), (iPTR imm)), + (AltTo.VT (!cast(NAME # To.EltSize # "x4rr") + VR512:$src1, + (EXTRACT_get_vextract_imm To.RC:$ext)))>; + + // A 128/256-bit subvector extract from the first 512-bit vector position is + // a subregister copy that needs no instruction. + def : Pat<(To.VT (extract_subvector (From.VT VR512:$src), (iPTR 0))), + (To.VT + (EXTRACT_SUBREG (From.VT VR512:$src), To.SubRegIdx))>; + + // And for the alternative types. + def : Pat<(AltTo.VT (extract_subvector (AltFrom.VT VR512:$src), (iPTR 0))), + (AltTo.VT + (EXTRACT_SUBREG (AltFrom.VT VR512:$src), AltTo.SubRegIdx))>; + + // Intrinsic call with masking. + def : Pat<(!cast("int_x86_avx512_mask_vextract" # To.EltTypeName # + "x4_512") + VR512:$src1, (iPTR imm:$idx), To.RC:$src0, GR8:$mask), + (!cast(NAME # To.EltSize # "x4rrk") To.RC:$src0, + (v4i1 (COPY_TO_REGCLASS GR8:$mask, VK4WM)), + VR512:$src1, imm:$idx)>; + + // Intrinsic call with zero-masking. + def : Pat<(!cast("int_x86_avx512_mask_vextract" # To.EltTypeName # + "x4_512") + VR512:$src1, (iPTR imm:$idx), To.ImmAllZerosV, GR8:$mask), + (!cast(NAME # To.EltSize # "x4rrkz") + (v4i1 (COPY_TO_REGCLASS GR8:$mask, VK4WM)), + VR512:$src1, imm:$idx)>; + + // Intrinsic call without masking. + def : Pat<(!cast("int_x86_avx512_mask_vextract" # To.EltTypeName # + "x4_512") + VR512:$src1, (iPTR imm:$idx), To.ImmAllZerosV, (i8 -1)), + (!cast(NAME # To.EltSize # "x4rr") + VR512:$src1, imm:$idx)>; } -let hasSideEffects = 0 in { -// -- 32x4 form -- -def VEXTRACTI32x4rr : AVX512AIi8<0x39, MRMDestReg, (outs VR128X:$dst), - (ins VR512:$src1, i8imm:$src2), - "vextracti32x4\t{$src2, $src1, $dst|$dst, $src1, $src2}", - []>, EVEX, EVEX_V512; -def VEXTRACTI32x4mr : AVX512AIi8<0x39, MRMDestMem, (outs), - (ins i128mem:$dst, VR512:$src1, i8imm:$src2), - "vextracti32x4\t{$src2, $src1, $dst|$dst, $src1, $src2}", - []>, EVEX, EVEX_V512, EVEX_CD8<32, CD8VT4>; - -// -- 64x4 form -- -def VEXTRACTI64x4rr : AVX512AIi8<0x3b, MRMDestReg, (outs VR256X:$dst), - (ins VR512:$src1, i8imm:$src2), - "vextracti64x4\t{$src2, $src1, $dst|$dst, $src1, $src2}", - []>, EVEX, EVEX_V512, VEX_W; -let mayStore = 1 in -def VEXTRACTI64x4mr : AVX512AIi8<0x3b, MRMDestMem, (outs), - (ins i256mem:$dst, VR512:$src1, i8imm:$src2), - "vextracti64x4\t{$src2, $src1, $dst|$dst, $src1, $src2}", - []>, EVEX, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VT4>; -} - -def : Pat<(vextract128_extract:$ext (v16f32 VR512:$src1), (iPTR imm)), - (v4f32 (VEXTRACTF32x4rr VR512:$src1, - (EXTRACT_get_vextract128_imm VR128X:$ext)))>; - -def : Pat<(vextract128_extract:$ext VR512:$src1, (iPTR imm)), - (v4i32 (VEXTRACTF32x4rr VR512:$src1, - (EXTRACT_get_vextract128_imm VR128X:$ext)))>; - -def : Pat<(vextract128_extract:$ext (v8f64 VR512:$src1), (iPTR imm)), - (v2f64 (VEXTRACTF32x4rr VR512:$src1, - (EXTRACT_get_vextract128_imm VR128X:$ext)))>; - -def : Pat<(vextract128_extract:$ext (v8i64 VR512:$src1), (iPTR imm)), - (v2i64 (VEXTRACTI32x4rr VR512:$src1, - (EXTRACT_get_vextract128_imm VR128X:$ext)))>; - - -def : Pat<(vextract256_extract:$ext (v16f32 VR512:$src1), (iPTR imm)), - (v8f32 (VEXTRACTF64x4rr VR512:$src1, - (EXTRACT_get_vextract256_imm VR256X:$ext)))>; - -def : Pat<(vextract256_extract:$ext (v16i32 VR512:$src1), (iPTR imm)), - (v8i32 (VEXTRACTI64x4rr VR512:$src1, - (EXTRACT_get_vextract256_imm VR256X:$ext)))>; - -def : Pat<(vextract256_extract:$ext (v8f64 VR512:$src1), (iPTR imm)), - (v4f64 (VEXTRACTF64x4rr VR512:$src1, - (EXTRACT_get_vextract256_imm VR256X:$ext)))>; - -def : Pat<(vextract256_extract:$ext (v8i64 VR512:$src1), (iPTR imm)), - (v4i64 (VEXTRACTI64x4rr VR512:$src1, - (EXTRACT_get_vextract256_imm VR256X:$ext)))>; - -// A 256-bit subvector extract from the first 512-bit vector position -// is a subregister copy that needs no instruction. -def : Pat<(v8i32 (extract_subvector (v16i32 VR512:$src), (iPTR 0))), - (v8i32 (EXTRACT_SUBREG (v16i32 VR512:$src), sub_ymm))>; -def : Pat<(v8f32 (extract_subvector (v16f32 VR512:$src), (iPTR 0))), - (v8f32 (EXTRACT_SUBREG (v16f32 VR512:$src), sub_ymm))>; -def : Pat<(v4i64 (extract_subvector (v8i64 VR512:$src), (iPTR 0))), - (v4i64 (EXTRACT_SUBREG (v8i64 VR512:$src), sub_ymm))>; -def : Pat<(v4f64 (extract_subvector (v8f64 VR512:$src), (iPTR 0))), - (v4f64 (EXTRACT_SUBREG (v8f64 VR512:$src), sub_ymm))>; - -// zmm -> xmm -def : Pat<(v4i32 (extract_subvector (v16i32 VR512:$src), (iPTR 0))), - (v4i32 (EXTRACT_SUBREG (v16i32 VR512:$src), sub_xmm))>; -def : Pat<(v2i64 (extract_subvector (v8i64 VR512:$src), (iPTR 0))), - (v2i64 (EXTRACT_SUBREG (v8i64 VR512:$src), sub_xmm))>; -def : Pat<(v2f64 (extract_subvector (v8f64 VR512:$src), (iPTR 0))), - (v2f64 (EXTRACT_SUBREG (v8f64 VR512:$src), sub_xmm))>; -def : Pat<(v4f32 (extract_subvector (v16f32 VR512:$src), (iPTR 0))), - (v4f32 (EXTRACT_SUBREG (v16f32 VR512:$src), sub_xmm))>; +multiclass vextract_for_type { + defm NAME # "32x4" : vextract_for_size, + X86VectorVTInfo< 4, EltVT32, VR128X>, + X86VectorVTInfo< 8, EltVT64, VR512>, + X86VectorVTInfo< 2, EltVT64, VR128X>, + vextract128_extract, + EXTRACT_get_vextract128_imm>; + defm NAME # "64x4" : vextract_for_size, + X86VectorVTInfo< 4, EltVT64, VR256X>, + X86VectorVTInfo<16, EltVT32, VR512>, + X86VectorVTInfo< 8, EltVT32, VR256>, + vextract256_extract, + EXTRACT_get_vextract256_imm>, VEX_W; +} +defm VEXTRACTF : vextract_for_type; +defm VEXTRACTI : vextract_for_type; // A 128-bit subvector insert to the first 512-bit vector position // is a subregister copy that needs no instruction. @@ -352,13 +555,13 @@ def : Pat<(insert_subvector undef, (v8f32 VR256X:$src), (iPTR 0)), // vextractps - extract 32 bits from XMM def VEXTRACTPSzrr : AVX512AIi8<0x17, MRMDestReg, (outs GR32:$dst), - (ins VR128X:$src1, u32u8imm:$src2), + (ins VR128X:$src1, i32i8imm:$src2), "vextractps\t{$src2, $src1, $dst|$dst, $src1, $src2}", [(set GR32:$dst, (extractelt (bc_v4i32 (v4f32 VR128X:$src1)), imm:$src2))]>, EVEX; def VEXTRACTPSzmr : AVX512AIi8<0x17, MRMDestMem, (outs), - (ins f32mem:$dst, VR128X:$src1, u32u8imm:$src2), + (ins f32mem:$dst, VR128X:$src1, i32i8imm:$src2), "vextractps\t{$src2, $src1, $dst|$dst, $src1, $src2}", [(store (extractelt (bc_v4i32 (v4f32 VR128X:$src1)), imm:$src2), addr:$dst)]>, EVEX, EVEX_CD8<32, CD8VT1>; @@ -476,6 +679,28 @@ defm VPBROADCASTQZ : avx512_int_broadcast_rm<0x59, "vpbroadcastq", i64mem, loadi64, VR512, v8i64, v2i64, VK8WM>, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VT1>; +multiclass avx512_int_subvec_broadcast_rm opc, string OpcodeStr, + X86MemOperand x86memop, PatFrag ld_frag, + RegisterClass KRC> { + let mayLoad = 1 in { + def rm : AVX5128I, EVEX; + def krm : AVX5128I, EVEX, EVEX_KZ; + } +} + +defm VBROADCASTI32X4 : avx512_int_subvec_broadcast_rm<0x5a, "vbroadcasti32x4", + i128mem, loadv2i64, VK16WM>, + EVEX_V512, EVEX_CD8<32, CD8VT4>; +defm VBROADCASTI64X4 : avx512_int_subvec_broadcast_rm<0x5b, "vbroadcasti64x4", + i256mem, loadv4i64, VK16WM>, VEX_W, + EVEX_V512, EVEX_CD8<64, CD8VT4>; + def : Pat<(v16i32 (int_x86_avx512_pbroadcastd_512 (v4i32 VR128X:$src))), (VPBROADCASTDZrr VR128X:$src)>; def : Pat<(v8i64 (int_x86_avx512_pbroadcastq_512 (v2i64 VR128X:$src))), @@ -510,46 +735,91 @@ def : Pat<(v8i32 (X86VBroadcastm (v8i1 VK8WM:$mask), (loadi32 addr:$src))), //--- multiclass avx512_mask_broadcast opc, string OpcodeStr, - RegisterClass DstRC, RegisterClass KRC, - ValueType OpVT, ValueType SrcVT> { -def rr : AVX512XS8I { +let Predicates = [HasCDI] in +def Zrr : AVX512XS8I, EVEX; + []>, EVEX, EVEX_V512; + +let Predicates = [HasCDI, HasVLX] in { +def Z128rr : AVX512XS8I, EVEX, EVEX_V128; +def Z256rr : AVX512XS8I, EVEX, EVEX_V256; +} } -defm VPBROADCASTMW2D : avx512_mask_broadcast<0x3A, "vpbroadcastmw2d", VR512, - VK16, v16i32, v16i1>, EVEX_V512; -defm VPBROADCASTMB2Q : avx512_mask_broadcast<0x2A, "vpbroadcastmb2q", VR512, - VK8, v8i64, v8i1>, EVEX_V512, VEX_W; +let Predicates = [HasCDI] in { +defm VPBROADCASTMW2D : avx512_mask_broadcast<0x3A, "vpbroadcastmw2d", + VK16>; +defm VPBROADCASTMB2Q : avx512_mask_broadcast<0x2A, "vpbroadcastmb2q", + VK8>, VEX_W; +} //===----------------------------------------------------------------------===// // AVX-512 - VPERM // // -- immediate form -- -multiclass avx512_perm_imm opc, string OpcodeStr, RegisterClass RC, - SDNode OpNode, PatFrag mem_frag, - X86MemOperand x86memop, ValueType OpVT> { - def ri : AVX512AIi8 opc, string OpcodeStr, SDNode OpNode, + X86VectorVTInfo _> { + let ExeDomain = _.ExeDomain in { + def ri : AVX512AIi8, + [(set _.RC:$dst, + (_.VT (OpNode _.RC:$src1, (i8 imm:$src2))))]>, EVEX; - def mi : AVX512AIi8, EVEX; + [(set _.RC:$dst, + (_.VT (OpNode (_.MemOpFrag addr:$src1), + (i8 imm:$src2))))]>, + EVEX, EVEX_CD8<_.EltSize, CD8VF>; +} +} + +multiclass avx512_permil OpcImm, bits<8> OpcVar, X86VectorVTInfo _, + X86VectorVTInfo Ctrl> : + avx512_perm_imm { + let ExeDomain = _.ExeDomain in { + def rr : AVX5128I, + EVEX_4V; + def rm : AVX5128I, + EVEX_4V; + } } -defm VPERMQZ : avx512_perm_imm<0x00, "vpermq", VR512, X86VPermi, memopv8i64, - i512mem, v8i64>, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>; -let ExeDomain = SSEPackedDouble in -defm VPERMPDZ : avx512_perm_imm<0x01, "vpermpd", VR512, X86VPermi, memopv8f64, - f512mem, v8f64>, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>; +defm VPERMQZ : avx512_perm_imm<0x00, "vpermq", X86VPermi, v8i64_info>, + EVEX_V512, VEX_W; +defm VPERMPDZ : avx512_perm_imm<0x01, "vpermpd", X86VPermi, v8f64_info>, + EVEX_V512, VEX_W; + +defm VPERMILPSZ : avx512_permil<0x04, 0x0C, v16f32_info, v16i32_info>, + EVEX_V512; +defm VPERMILPDZ : avx512_permil<0x05, 0x0D, v8f64_info, v8i64_info>, + EVEX_V512, VEX_W; + +def : Pat<(v16i32 (X86VPermilpi VR512:$src1, (i8 imm:$imm))), + (VPERMILPSZri VR512:$src1, imm:$imm)>; +def : Pat<(v8i64 (X86VPermilpi VR512:$src1, (i8 imm:$imm))), + (VPERMILPDZri VR512:$src1, imm:$imm)>; // -- VPERM - register form -- multiclass avx512_perm opc, string OpcodeStr, RegisterClass RC, @@ -585,7 +855,7 @@ defm VPERMPDZ : avx512_perm<0x16, "vpermpd", VR512, memopv8f64, f512mem, // -- VPERM2I - 3 source operands form -- multiclass avx512_perm_3src opc, string OpcodeStr, RegisterClass RC, PatFrag mem_frag, X86MemOperand x86memop, - SDNode OpNode, ValueType OpVT> { + SDNode OpNode, ValueType OpVT, RegisterClass KRC> { let Constraints = "$src1 = $dst" in { def rr : AVX5128I, EVEX_4V; + def rrk : AVX5128I, + EVEX_4V, EVEX_K; + + let AddedComplexity = 30 in // Prefer over VMOV*rrkz Pat<> + def rrkz : AVX5128I, + EVEX_4V, EVEX_KZ; + def rm : AVX5128I, EVEX_4V; + + def rmk : AVX5128I, + EVEX_4V, EVEX_K; + + let AddedComplexity = 10 in // Prefer over the rrkz variant + def rmkz : AVX5128I, + EVEX_4V, EVEX_KZ; } } -defm VPERMI2D : avx512_perm_3src<0x76, "vpermi2d", VR512, memopv16i32, i512mem, - X86VPermiv3, v16i32>, EVEX_V512, EVEX_CD8<32, CD8VF>; -defm VPERMI2Q : avx512_perm_3src<0x76, "vpermi2q", VR512, memopv8i64, i512mem, - X86VPermiv3, v8i64>, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>; -defm VPERMI2PS : avx512_perm_3src<0x77, "vpermi2ps", VR512, memopv16f32, i512mem, - X86VPermiv3, v16f32>, EVEX_V512, EVEX_CD8<32, CD8VF>; -defm VPERMI2PD : avx512_perm_3src<0x77, "vpermi2pd", VR512, memopv8f64, i512mem, - X86VPermiv3, v8f64>, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>; - -defm VPERMT2D : avx512_perm_3src<0x7E, "vpermt2d", VR512, memopv16i32, i512mem, - X86VPermv3, v16i32>, EVEX_V512, EVEX_CD8<32, CD8VF>; -defm VPERMT2Q : avx512_perm_3src<0x7E, "vpermt2q", VR512, memopv8i64, i512mem, - X86VPermv3, v8i64>, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>; -defm VPERMT2PS : avx512_perm_3src<0x7F, "vpermt2ps", VR512, memopv16f32, i512mem, - X86VPermv3, v16f32>, EVEX_V512, EVEX_CD8<32, CD8VF>; -defm VPERMT2PD : avx512_perm_3src<0x7F, "vpermt2pd", VR512, memopv8f64, i512mem, - X86VPermv3, v8f64>, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>; +defm VPERMI2D : avx512_perm_3src<0x76, "vpermi2d", VR512, memopv16i32, + i512mem, X86VPermiv3, v16i32, VK16WM>, + EVEX_V512, EVEX_CD8<32, CD8VF>; +defm VPERMI2Q : avx512_perm_3src<0x76, "vpermi2q", VR512, memopv8i64, + i512mem, X86VPermiv3, v8i64, VK8WM>, + EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>; +defm VPERMI2PS : avx512_perm_3src<0x77, "vpermi2ps", VR512, memopv16f32, + i512mem, X86VPermiv3, v16f32, VK16WM>, + EVEX_V512, EVEX_CD8<32, CD8VF>; +defm VPERMI2PD : avx512_perm_3src<0x77, "vpermi2pd", VR512, memopv8f64, + i512mem, X86VPermiv3, v8f64, VK8WM>, + EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>; + +multiclass avx512_perm_table_3src opc, string Suffix, RegisterClass RC, + PatFrag mem_frag, X86MemOperand x86memop, + SDNode OpNode, ValueType OpVT, RegisterClass KRC, + ValueType MaskVT, RegisterClass MRC> : + avx512_perm_3src { + def : Pat<(OpVT (!cast("int_x86_avx512_mask_vpermt_"##Suffix##"_512") + VR512:$idx, VR512:$src1, VR512:$src2, -1)), + (!cast(NAME#rr) VR512:$src1, VR512:$idx, VR512:$src2)>; + + def : Pat<(OpVT (!cast("int_x86_avx512_mask_vpermt_"##Suffix##"_512") + VR512:$idx, VR512:$src1, VR512:$src2, MRC:$mask)), + (!cast(NAME#rrk) VR512:$src1, + (MaskVT (COPY_TO_REGCLASS MRC:$mask, KRC)), VR512:$idx, VR512:$src2)>; +} + +defm VPERMT2D : avx512_perm_table_3src<0x7E, "d", VR512, memopv16i32, i512mem, + X86VPermv3, v16i32, VK16WM, v16i1, GR16>, + EVEX_V512, EVEX_CD8<32, CD8VF>; +defm VPERMT2Q : avx512_perm_table_3src<0x7E, "q", VR512, memopv8i64, i512mem, + X86VPermv3, v8i64, VK8WM, v8i1, GR8>, + EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>; +defm VPERMT2PS : avx512_perm_table_3src<0x7F, "ps", VR512, memopv16f32, i512mem, + X86VPermv3, v16f32, VK16WM, v16i1, GR16>, + EVEX_V512, EVEX_CD8<32, CD8VF>; +defm VPERMT2PD : avx512_perm_table_3src<0x7F, "pd", VR512, memopv8f64, i512mem, + X86VPermv3, v8f64, VK8WM, v8i1, GR8>, + EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>; + //===----------------------------------------------------------------------===// // AVX-512 - BLEND using mask // @@ -735,91 +1080,297 @@ defm VCMPSDZ : avx512_cmp_scalar opc, string OpcodeStr, RegisterClass KRC, - RegisterClass RC, X86MemOperand x86memop, PatFrag memop_frag, - SDNode OpNode, ValueType vt> { +multiclass avx512_icmp_packed opc, string OpcodeStr, SDNode OpNode, + X86VectorVTInfo _> { def rr : AVX512BI, EVEX_4V; + let mayLoad = 1 in def rm : AVX512BI, EVEX_4V; + def rrk : AVX512BI, EVEX_4V, EVEX_K; + let mayLoad = 1 in + def rmk : AVX512BI, EVEX_4V, EVEX_K; } -defm VPCMPEQDZ : avx512_icmp_packed<0x76, "vpcmpeqd", VK16, VR512, i512mem, - memopv16i32, X86pcmpeqm, v16i32>, EVEX_V512, - EVEX_CD8<32, CD8VF>; -defm VPCMPEQQZ : avx512_icmp_packed<0x29, "vpcmpeqq", VK8, VR512, i512mem, - memopv8i64, X86pcmpeqm, v8i64>, T8PD, EVEX_V512, - VEX_W, EVEX_CD8<64, CD8VF>; +multiclass avx512_icmp_packed_rmb opc, string OpcodeStr, SDNode OpNode, + X86VectorVTInfo _> : + avx512_icmp_packed { + let mayLoad = 1 in { + def rmb : AVX512BI, EVEX_4V, EVEX_B; + def rmbk : AVX512BI, EVEX_4V, EVEX_K, EVEX_B; + } +} -defm VPCMPGTDZ : avx512_icmp_packed<0x66, "vpcmpgtd", VK16, VR512, i512mem, - memopv16i32, X86pcmpgtm, v16i32>, EVEX_V512, - EVEX_CD8<32, CD8VF>; -defm VPCMPGTQZ : avx512_icmp_packed<0x37, "vpcmpgtq", VK8, VR512, i512mem, - memopv8i64, X86pcmpgtm, v8i64>, T8PD, EVEX_V512, - VEX_W, EVEX_CD8<64, CD8VF>; +multiclass avx512_icmp_packed_vl opc, string OpcodeStr, SDNode OpNode, + AVX512VLVectorVTInfo VTInfo, Predicate prd> { + let Predicates = [prd] in + defm Z : avx512_icmp_packed, + EVEX_V512; + + let Predicates = [prd, HasVLX] in { + defm Z256 : avx512_icmp_packed, + EVEX_V256; + defm Z128 : avx512_icmp_packed, + EVEX_V128; + } +} + +multiclass avx512_icmp_packed_rmb_vl opc, string OpcodeStr, + SDNode OpNode, AVX512VLVectorVTInfo VTInfo, + Predicate prd> { + let Predicates = [prd] in + defm Z : avx512_icmp_packed_rmb, + EVEX_V512; + + let Predicates = [prd, HasVLX] in { + defm Z256 : avx512_icmp_packed_rmb, + EVEX_V256; + defm Z128 : avx512_icmp_packed_rmb, + EVEX_V128; + } +} + +defm VPCMPEQB : avx512_icmp_packed_vl<0x74, "vpcmpeqb", X86pcmpeqm, + avx512vl_i8_info, HasBWI>, + EVEX_CD8<8, CD8VF>; + +defm VPCMPEQW : avx512_icmp_packed_vl<0x75, "vpcmpeqw", X86pcmpeqm, + avx512vl_i16_info, HasBWI>, + EVEX_CD8<16, CD8VF>; + +defm VPCMPEQD : avx512_icmp_packed_rmb_vl<0x76, "vpcmpeqd", X86pcmpeqm, + avx512vl_i32_info, HasAVX512>, + EVEX_CD8<32, CD8VF>; + +defm VPCMPEQQ : avx512_icmp_packed_rmb_vl<0x29, "vpcmpeqq", X86pcmpeqm, + avx512vl_i64_info, HasAVX512>, + T8PD, VEX_W, EVEX_CD8<64, CD8VF>; + +defm VPCMPGTB : avx512_icmp_packed_vl<0x64, "vpcmpgtb", X86pcmpgtm, + avx512vl_i8_info, HasBWI>, + EVEX_CD8<8, CD8VF>; + +defm VPCMPGTW : avx512_icmp_packed_vl<0x65, "vpcmpgtw", X86pcmpgtm, + avx512vl_i16_info, HasBWI>, + EVEX_CD8<16, CD8VF>; + +defm VPCMPGTD : avx512_icmp_packed_rmb_vl<0x66, "vpcmpgtd", X86pcmpgtm, + avx512vl_i32_info, HasAVX512>, + EVEX_CD8<32, CD8VF>; + +defm VPCMPGTQ : avx512_icmp_packed_rmb_vl<0x37, "vpcmpgtq", X86pcmpgtm, + avx512vl_i64_info, HasAVX512>, + T8PD, VEX_W, EVEX_CD8<64, CD8VF>; def : Pat<(v8i1 (X86pcmpgtm (v8i32 VR256X:$src1), (v8i32 VR256X:$src2))), - (COPY_TO_REGCLASS (VPCMPGTDZrr + (COPY_TO_REGCLASS (VPCMPGTDZrr (v16i32 (SUBREG_TO_REG (i32 0), VR256X:$src1, sub_ymm)), (v16i32 (SUBREG_TO_REG (i32 0), VR256X:$src2, sub_ymm))), VK8)>; def : Pat<(v8i1 (X86pcmpeqm (v8i32 VR256X:$src1), (v8i32 VR256X:$src2))), - (COPY_TO_REGCLASS (VPCMPEQDZrr + (COPY_TO_REGCLASS (VPCMPEQDZrr (v16i32 (SUBREG_TO_REG (i32 0), VR256X:$src1, sub_ymm)), (v16i32 (SUBREG_TO_REG (i32 0), VR256X:$src2, sub_ymm))), VK8)>; -multiclass avx512_icmp_cc opc, RegisterClass KRC, - RegisterClass RC, X86MemOperand x86memop, PatFrag memop_frag, - SDNode OpNode, ValueType vt, Operand CC, string asm, - string asm_alt> { +multiclass avx512_icmp_cc opc, string Suffix, SDNode OpNode, + X86VectorVTInfo _> { def rri : AVX512AIi8, EVEX_4V; + let mayLoad = 1 in def rmi : AVX512AIi8, EVEX_4V; + (outs _.KRC:$dst), (ins _.RC:$src1, _.MemOp:$src2, AVXCC:$cc), + !strconcat("vpcmp${cc}", Suffix, + "\t{$src2, $src1, $dst|$dst, $src1, $src2}"), + [(set _.KRC:$dst, (OpNode (_.VT _.RC:$src1), + (_.VT (bitconvert (_.LdFrag addr:$src2))), + imm:$cc))], + IIC_SSE_ALU_F32P_RM>, EVEX_4V; + def rrik : AVX512AIi8, EVEX_4V, EVEX_K; + let mayLoad = 1 in + def rmik : AVX512AIi8, EVEX_4V, EVEX_K; + // Accept explicit immediate argument form instead of comparison code. let isAsmParserOnly = 1, hasSideEffects = 0 in { def rri_alt : AVX512AIi8, EVEX_4V; + (outs _.KRC:$dst), (ins _.RC:$src1, _.RC:$src2, i8imm:$cc), + !strconcat("vpcmp", Suffix, "\t{$cc, $src2, $src1, $dst|", + "$dst, $src1, $src2, $cc}"), + [], IIC_SSE_ALU_F32P_RR>, EVEX_4V; def rmi_alt : AVX512AIi8, EVEX_4V; + (outs _.KRC:$dst), (ins _.RC:$src1, _.MemOp:$src2, i8imm:$cc), + !strconcat("vpcmp", Suffix, "\t{$cc, $src2, $src1, $dst|", + "$dst, $src1, $src2, $cc}"), + [], IIC_SSE_ALU_F32P_RM>, EVEX_4V; + def rrik_alt : AVX512AIi8, EVEX_4V, EVEX_K; + def rmik_alt : AVX512AIi8, EVEX_4V, EVEX_K; + } +} + +multiclass avx512_icmp_cc_rmb opc, string Suffix, SDNode OpNode, + X86VectorVTInfo _> : + avx512_icmp_cc { + let mayLoad = 1 in { + def rmib : AVX512AIi8, EVEX_4V, EVEX_B; + def rmibk : AVX512AIi8, EVEX_4V, EVEX_K, EVEX_B; + } + + // Accept explicit immediate argument form instead of comparison code. + let isAsmParserOnly = 1, hasSideEffects = 0 in { + def rmib_alt : AVX512AIi8, EVEX_4V, EVEX_B; + def rmibk_alt : AVX512AIi8, EVEX_4V, EVEX_K, EVEX_B; } } -defm VPCMPDZ : avx512_icmp_cc<0x1F, VK16, VR512, i512mem, memopv16i32, - X86cmpm, v16i32, AVXCC, - "vpcmp${cc}d\t{$src2, $src1, $dst|$dst, $src1, $src2}", - "vpcmpd\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}">, - EVEX_V512, EVEX_CD8<32, CD8VF>; -defm VPCMPUDZ : avx512_icmp_cc<0x1E, VK16, VR512, i512mem, memopv16i32, - X86cmpmu, v16i32, AVXCC, - "vpcmp${cc}ud\t{$src2, $src1, $dst|$dst, $src1, $src2}", - "vpcmpud\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}">, - EVEX_V512, EVEX_CD8<32, CD8VF>; - -defm VPCMPQZ : avx512_icmp_cc<0x1F, VK8, VR512, i512mem, memopv8i64, - X86cmpm, v8i64, AVXCC, - "vpcmp${cc}q\t{$src2, $src1, $dst|$dst, $src1, $src2}", - "vpcmpq\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}">, - VEX_W, EVEX_V512, EVEX_CD8<64, CD8VF>; -defm VPCMPUQZ : avx512_icmp_cc<0x1E, VK8, VR512, i512mem, memopv8i64, - X86cmpmu, v8i64, AVXCC, - "vpcmp${cc}uq\t{$src2, $src1, $dst|$dst, $src1, $src2}", - "vpcmpuq\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}">, - VEX_W, EVEX_V512, EVEX_CD8<64, CD8VF>; - -// avx512_cmp_packed - sse 1 & 2 compare packed instructions +multiclass avx512_icmp_cc_vl opc, string Suffix, SDNode OpNode, + AVX512VLVectorVTInfo VTInfo, Predicate prd> { + let Predicates = [prd] in + defm Z : avx512_icmp_cc, EVEX_V512; + + let Predicates = [prd, HasVLX] in { + defm Z256 : avx512_icmp_cc, EVEX_V256; + defm Z128 : avx512_icmp_cc, EVEX_V128; + } +} + +multiclass avx512_icmp_cc_rmb_vl opc, string Suffix, SDNode OpNode, + AVX512VLVectorVTInfo VTInfo, Predicate prd> { + let Predicates = [prd] in + defm Z : avx512_icmp_cc_rmb, + EVEX_V512; + + let Predicates = [prd, HasVLX] in { + defm Z256 : avx512_icmp_cc_rmb, + EVEX_V256; + defm Z128 : avx512_icmp_cc_rmb, + EVEX_V128; + } +} + +defm VPCMPB : avx512_icmp_cc_vl<0x3F, "b", X86cmpm, avx512vl_i8_info, + HasBWI>, EVEX_CD8<8, CD8VF>; +defm VPCMPUB : avx512_icmp_cc_vl<0x3E, "ub", X86cmpmu, avx512vl_i8_info, + HasBWI>, EVEX_CD8<8, CD8VF>; + +defm VPCMPW : avx512_icmp_cc_vl<0x3F, "w", X86cmpm, avx512vl_i16_info, + HasBWI>, VEX_W, EVEX_CD8<16, CD8VF>; +defm VPCMPUW : avx512_icmp_cc_vl<0x3E, "uw", X86cmpmu, avx512vl_i16_info, + HasBWI>, VEX_W, EVEX_CD8<16, CD8VF>; + +defm VPCMPD : avx512_icmp_cc_rmb_vl<0x1F, "d", X86cmpm, avx512vl_i32_info, + HasAVX512>, EVEX_CD8<32, CD8VF>; +defm VPCMPUD : avx512_icmp_cc_rmb_vl<0x1E, "ud", X86cmpmu, avx512vl_i32_info, + HasAVX512>, EVEX_CD8<32, CD8VF>; + +defm VPCMPQ : avx512_icmp_cc_rmb_vl<0x1F, "q", X86cmpm, avx512vl_i64_info, + HasAVX512>, VEX_W, EVEX_CD8<64, CD8VF>; +defm VPCMPUQ : avx512_icmp_cc_rmb_vl<0x1E, "uq", X86cmpmu, avx512vl_i64_info, + HasAVX512>, VEX_W, EVEX_CD8<64, CD8VF>; + +// avx512_cmp_packed - compare packed instructions multiclass avx512_cmp_packed { @@ -843,11 +1394,11 @@ multiclass avx512_cmp_packed; def rmi_alt : AVX512PIi8<0xC2, MRMSrcMem, - (outs RC:$dst), (ins RC:$src1, x86memop:$src2, i8imm:$cc), + (outs KRC:$dst), (ins RC:$src1, x86memop:$src2, i8imm:$cc), !strconcat("vcmp", suffix, " \t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}"), [], d>; } @@ -907,14 +1458,14 @@ def : Pat<(i8 (int_x86_avx512_mask_cmp_pd_512 (v8f64 VR512:$src1), // multiclass avx512_mask_mov opc_kk, bits<8> opc_km, bits<8> opc_mk, string OpcodeStr, RegisterClass KRC, - ValueType vt, X86MemOperand x86memop> { + ValueType vvt, ValueType ivt, X86MemOperand x86memop> { let hasSideEffects = 0 in { def kk : I; let mayLoad = 1 in def km : I; + [(set KRC:$dst, (vvt (bitconvert (ivt (load addr:$src)))))]>; let mayStore = 1 in def mk : I; @@ -932,32 +1483,82 @@ multiclass avx512_mask_mov_gpr opc_kr, bits<8> opc_rk, } } -let Predicates = [HasAVX512] in { - defm KMOVW : avx512_mask_mov<0x90, 0x90, 0x91, "kmovw", VK16, v16i1, i16mem>, - VEX, PS; - defm KMOVW : avx512_mask_mov_gpr<0x92, 0x93, "kmovw", VK16, GR32>, +let Predicates = [HasDQI] in + defm KMOVB : avx512_mask_mov<0x90, 0x90, 0x91, "kmovb", VK8, v8i1, i8, + i8mem>, + avx512_mask_mov_gpr<0x92, 0x93, "kmovb", VK8, GR32>, + VEX, PD; + +let Predicates = [HasAVX512] in + defm KMOVW : avx512_mask_mov<0x90, 0x90, 0x91, "kmovw", VK16, v16i1, i16, + i16mem>, + avx512_mask_mov_gpr<0x92, 0x93, "kmovw", VK16, GR32>, VEX, PS; + +let Predicates = [HasBWI] in { + defm KMOVD : avx512_mask_mov<0x90, 0x90, 0x91, "kmovd", VK32, v32i1, i32, + i32mem>, VEX, PD, VEX_W; + defm KMOVD : avx512_mask_mov_gpr<0x92, 0x93, "kmovd", VK32, GR32>, + VEX, XD; +} + +let Predicates = [HasBWI] in { + defm KMOVQ : avx512_mask_mov<0x90, 0x90, 0x91, "kmovq", VK64, v64i1, i64, + i64mem>, VEX, PS, VEX_W; + defm KMOVQ : avx512_mask_mov_gpr<0x92, 0x93, "kmovq", VK64, GR64>, + VEX, XD, VEX_W; } +// GR from/to mask register +let Predicates = [HasDQI] in { + def : Pat<(v8i1 (bitconvert (i8 GR8:$src))), + (KMOVBkr (SUBREG_TO_REG (i32 0), GR8:$src, sub_8bit))>; + def : Pat<(i8 (bitconvert (v8i1 VK8:$src))), + (EXTRACT_SUBREG (KMOVBrk VK8:$src), sub_8bit)>; +} let Predicates = [HasAVX512] in { - // GR16 from/to 16-bit mask def : Pat<(v16i1 (bitconvert (i16 GR16:$src))), (KMOVWkr (SUBREG_TO_REG (i32 0), GR16:$src, sub_16bit))>; def : Pat<(i16 (bitconvert (v16i1 VK16:$src))), (EXTRACT_SUBREG (KMOVWrk VK16:$src), sub_16bit)>; +} +let Predicates = [HasBWI] in { + def : Pat<(v32i1 (bitconvert (i32 GR32:$src))), (KMOVDkr GR32:$src)>; + def : Pat<(i32 (bitconvert (v32i1 VK32:$src))), (KMOVDrk VK32:$src)>; +} +let Predicates = [HasBWI] in { + def : Pat<(v64i1 (bitconvert (i64 GR64:$src))), (KMOVQkr GR64:$src)>; + def : Pat<(i64 (bitconvert (v64i1 VK64:$src))), (KMOVQrk VK64:$src)>; +} - // Store kreg in memory - def : Pat<(store (v16i1 VK16:$src), addr:$dst), +// Load/store kreg +let Predicates = [HasDQI] in { + def : Pat<(store (i8 (bitconvert (v8i1 VK8:$src))), addr:$dst), + (KMOVBmk addr:$dst, VK8:$src)>; +} +let Predicates = [HasAVX512] in { + def : Pat<(store (i16 (bitconvert (v16i1 VK16:$src))), addr:$dst), (KMOVWmk addr:$dst, VK16:$src)>; - - def : Pat<(store VK8:$src, addr:$dst), + def : Pat<(store (i8 (bitconvert (v8i1 VK8:$src))), addr:$dst), (KMOVWmk addr:$dst, (COPY_TO_REGCLASS VK8:$src, VK16))>; - def : Pat<(i1 (load addr:$src)), (COPY_TO_REGCLASS (KMOVWkm addr:$src), VK1)>; - - def : Pat<(v8i1 (load addr:$src)), + def : Pat<(v8i1 (bitconvert (i8 (load addr:$src)))), (COPY_TO_REGCLASS (KMOVWkm addr:$src), VK8)>; +} +let Predicates = [HasBWI] in { + def : Pat<(store (i32 (bitconvert (v32i1 VK32:$src))), addr:$dst), + (KMOVDmk addr:$dst, VK32:$src)>; +} +let Predicates = [HasBWI] in { + def : Pat<(store (i64 (bitconvert (v64i1 VK64:$src))), addr:$dst), + (KMOVQmk addr:$dst, VK64:$src)>; +} + +let Predicates = [HasAVX512] in { + def : Pat<(i1 (trunc (i64 GR64:$src))), + (COPY_TO_REGCLASS (KMOVWkr (AND32ri (EXTRACT_SUBREG $src, sub_32bit), + (i32 1))), VK1)>; def : Pat<(i1 (trunc (i32 GR32:$src))), (COPY_TO_REGCLASS (KMOVWkr (AND32ri $src, (i32 1))), VK1)>; @@ -970,7 +1571,7 @@ let Predicates = [HasAVX512] in { (COPY_TO_REGCLASS (KMOVWkr (AND32ri (SUBREG_TO_REG (i32 0), $src, sub_16bit), (i32 1))), VK1)>; - + def : Pat<(i32 (zext VK1:$src)), (AND32ri (KMOVWrk (COPY_TO_REGCLASS VK1:$src, VK16)), (i32 1))>; def : Pat<(i8 (zext VK1:$src)), @@ -989,6 +1590,14 @@ let Predicates = [HasAVX512] in { 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)>; +} + + // With AVX-512 only, 8-bit mask is promoted to 16-bit mask. let Predicates = [HasAVX512] in { // GR from/to 8-bit mask without native support @@ -1005,26 +1614,38 @@ let Predicates = [HasAVX512] in { (COPY_TO_REGCLASS VK16:$src, VK1)>; def : Pat<(i1 (X86Vextract VK8:$src, (iPTR 0))), (COPY_TO_REGCLASS VK8:$src, VK1)>; - +} +let Predicates = [HasBWI] in { + def : Pat<(i1 (X86Vextract VK32:$src, (iPTR 0))), + (COPY_TO_REGCLASS VK32:$src, VK1)>; + def : Pat<(i1 (X86Vextract VK64:$src, (iPTR 0))), + (COPY_TO_REGCLASS VK64:$src, VK1)>; } // Mask unary operation // - KNOT multiclass avx512_mask_unop opc, string OpcodeStr, - RegisterClass KRC, SDPatternOperator OpNode> { - let Predicates = [HasAVX512] in + RegisterClass KRC, SDPatternOperator OpNode, + Predicate prd> { + let Predicates = [prd] in def rr : I; } -multiclass avx512_mask_unop_w opc, string OpcodeStr, - SDPatternOperator OpNode> { - defm W : avx512_mask_unop, - VEX, PS; +multiclass avx512_mask_unop_all opc, string OpcodeStr, + SDPatternOperator OpNode> { + defm B : avx512_mask_unop, VEX, PD; + defm W : avx512_mask_unop, VEX, PS; + defm D : avx512_mask_unop, VEX, PD, VEX_W; + defm Q : avx512_mask_unop, VEX, PS, VEX_W; } -defm KNOT : avx512_mask_unop_w<0x44, "knot", not>; +defm KNOT : avx512_mask_unop_all<0x44, "knot", not>; multiclass avx512_mask_unop_int { let Predicates = [HasAVX512] in @@ -1035,43 +1656,60 @@ multiclass avx512_mask_unop_int { } defm : avx512_mask_unop_int<"knot", "KNOT">; +let Predicates = [HasDQI] in +def : Pat<(xor VK8:$src1, (v8i1 immAllOnesV)), (KNOTBrr VK8:$src1)>; +let Predicates = [HasAVX512] in def : Pat<(xor VK16:$src1, (v16i1 immAllOnesV)), (KNOTWrr VK16:$src1)>; +let Predicates = [HasBWI] in +def : Pat<(xor VK32:$src1, (v32i1 immAllOnesV)), (KNOTDrr VK32:$src1)>; +let Predicates = [HasBWI] in +def : Pat<(xor VK64:$src1, (v64i1 immAllOnesV)), (KNOTQrr VK64:$src1)>; + +// KNL does not support KMOVB, 8-bit mask is promoted to 16-bit +let Predicates = [HasAVX512] in { def : Pat<(xor VK8:$src1, (v8i1 immAllOnesV)), (COPY_TO_REGCLASS (KNOTWrr (COPY_TO_REGCLASS VK8:$src1, VK16)), VK8)>; -// With AVX-512, 8-bit mask is promoted to 16-bit mask. def : Pat<(not VK8:$src), (COPY_TO_REGCLASS (KNOTWrr (COPY_TO_REGCLASS VK8:$src, VK16)), VK8)>; +} // Mask binary operation // - KAND, KANDN, KOR, KXNOR, KXOR multiclass avx512_mask_binop opc, string OpcodeStr, - RegisterClass KRC, SDPatternOperator OpNode> { - let Predicates = [HasAVX512] in + RegisterClass KRC, SDPatternOperator OpNode, + Predicate prd> { + let Predicates = [prd] in def rr : I; } -multiclass avx512_mask_binop_w opc, string OpcodeStr, - SDPatternOperator OpNode> { - defm W : avx512_mask_binop, - VEX_4V, VEX_L, PS; +multiclass avx512_mask_binop_all opc, string OpcodeStr, + SDPatternOperator OpNode> { + defm B : avx512_mask_binop, VEX_4V, VEX_L, PD; + defm W : avx512_mask_binop, VEX_4V, VEX_L, PS; + defm D : avx512_mask_binop, VEX_4V, VEX_L, VEX_W, PD; + defm Q : avx512_mask_binop, VEX_4V, VEX_L, VEX_W, PS; } def andn : PatFrag<(ops node:$i0, node:$i1), (and (not node:$i0), node:$i1)>; def xnor : PatFrag<(ops node:$i0, node:$i1), (not (xor node:$i0, node:$i1))>; let isCommutable = 1 in { - defm KAND : avx512_mask_binop_w<0x41, "kand", and>; - let isCommutable = 0 in - defm KANDN : avx512_mask_binop_w<0x42, "kandn", andn>; - defm KOR : avx512_mask_binop_w<0x45, "kor", or>; - defm KXNOR : avx512_mask_binop_w<0x46, "kxnor", xnor>; - defm KXOR : avx512_mask_binop_w<0x47, "kxor", xor>; + defm KAND : avx512_mask_binop_all<0x41, "kand", and>; + defm KOR : avx512_mask_binop_all<0x45, "kor", or>; + defm KXNOR : avx512_mask_binop_all<0x46, "kxnor", xnor>; + defm KXOR : avx512_mask_binop_all<0x47, "kxor", xor>; } +let isCommutable = 0 in + defm KANDN : avx512_mask_binop_all<0x42, "kandn", andn>; def : Pat<(xor VK1:$src1, VK1:$src2), (COPY_TO_REGCLASS (KXORWrr (COPY_TO_REGCLASS VK1:$src1, VK16), @@ -1217,6 +1855,17 @@ 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:$src, (i8 8)), VK8))>; +let Predicates = [HasVLX] in { + def : Pat<(v8i1 (insert_subvector undef, (v4i1 VK4:$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:$src, VK8))>; + def : Pat<(v4i1 (extract_subvector (v8i1 VK8:$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:$src, VK2))>; +} + def : Pat<(v8i1 (X86vshli VK8:$src, (i8 imm:$imm))), (v8i1 (COPY_TO_REGCLASS (KSHIFTLWri (COPY_TO_REGCLASS VK8:$src, VK16), (I8Imm $imm)), VK8))>; @@ -1226,104 +1875,176 @@ def : Pat<(v8i1 (X86vsrli VK8:$src, (i8 imm:$imm))), // AVX-512 - Aligned and unaligned load and store // -multiclass avx512_load opc, RegisterClass RC, RegisterClass KRC, - X86MemOperand x86memop, PatFrag ld_frag, - string asm, Domain d, - ValueType vt, bit IsReMaterializable = 1> { +multiclass avx512_load opc, string OpcodeStr, PatFrag ld_frag, + RegisterClass KRC, RegisterClass RC, + ValueType vt, ValueType zvt, X86MemOperand memop, + Domain d, bit IsReMaterializable = 1> { let hasSideEffects = 0 in { def rr : AVX512PI, - EVEX; + !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"), [], + d>, EVEX; def rrkz : AVX512PI, EVEX, EVEX_KZ; + !strconcat(OpcodeStr, "\t{$src, ${dst} {${mask}} {z}|", + "${dst} {${mask}} {z}, $src}"), [], d>, EVEX, EVEX_KZ; } - let canFoldAsLoad = 1, isReMaterializable = IsReMaterializable in - def rm : AVX512PI, EVEX; - let Constraints = "$src1 = $dst", hasSideEffects = 0 in { - def rrk : AVX512PI, - EVEX, EVEX_K; - let mayLoad = 1 in - def rmk : AVX512PI, EVEX, EVEX_K; + let canFoldAsLoad = 1, isReMaterializable = IsReMaterializable, + SchedRW = [WriteLoad] in + def rm : AVX512PI, EVEX; + + let AddedComplexity = 20 in { + let Constraints = "$src0 = $dst", hasSideEffects = 0 in { + let hasSideEffects = 0 in + def rrk : AVX512PI, EVEX, EVEX_K; + let mayLoad = 1, SchedRW = [WriteLoad] in + def rmk : AVX512PI, EVEX, EVEX_K; + } + let mayLoad = 1, SchedRW = [WriteLoad] in + def rmkz : AVX512PI, EVEX, EVEX_KZ; + } +} + +multiclass avx512_load_vl opc, string OpcodeStr, string ld_pat, + string elty, string elsz, string vsz512, + string vsz256, string vsz128, Domain d, + Predicate prd, bit IsReMaterializable = 1> { + let Predicates = [prd] in + defm Z : avx512_load(ld_pat##"v"##vsz512##elty##elsz), + !cast("VK"##vsz512##"WM"), VR512, + !cast("v"##vsz512##elty##elsz), v16i32, + !cast(elty##"512mem"), d, + IsReMaterializable>, EVEX_V512; + + let Predicates = [prd, HasVLX] in { + defm Z256 : avx512_load(ld_pat##!if(!eq(elty,"f"), + "v"##vsz256##elty##elsz, "v4i64")), + !cast("VK"##vsz256##"WM"), VR256X, + !cast("v"##vsz256##elty##elsz), v8i32, + !cast(elty##"256mem"), d, + IsReMaterializable>, EVEX_V256; + + defm Z128 : avx512_load(ld_pat##!if(!eq(elty,"f"), + "v"##vsz128##elty##elsz, "v2i64")), + !cast("VK"##vsz128##"WM"), VR128X, + !cast("v"##vsz128##elty##elsz), v4i32, + !cast(elty##"128mem"), d, + IsReMaterializable>, EVEX_V128; } - let mayLoad = 1 in - def rmkz : AVX512PI, EVEX, EVEX_KZ; } -multiclass avx512_store opc, RegisterClass RC, RegisterClass KRC, - X86MemOperand x86memop, PatFrag store_frag, - string asm, Domain d, ValueType vt> { + +multiclass avx512_store opc, string OpcodeStr, PatFrag st_frag, + ValueType OpVT, RegisterClass KRC, RegisterClass RC, + X86MemOperand memop, Domain d> { let isAsmParserOnly = 1, hasSideEffects = 0 in { def rr_alt : AVX512PI, + !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"), [], d>, EVEX; let Constraints = "$src1 = $dst" in - def alt_rrk : AVX512PI, + def rrk_alt : AVX512PI, EVEX, EVEX_K; - def alt_rrkz : AVX512PI, EVEX, EVEX_KZ; } let mayStore = 1 in { - def mr : AVX512PI, EVEX; + def mr : AVX512PI, EVEX; def mrk : AVX512PI, EVEX, EVEX_K; - def mrkz : AVX512PI, EVEX, EVEX_KZ; } } -defm VMOVAPSZ : avx512_load<0x28, VR512, VK16WM, f512mem, alignedloadv16f32, - "vmovaps", SSEPackedSingle, v16f32>, - avx512_store<0x29, VR512, VK16WM, f512mem, alignedstore512, - "vmovaps", SSEPackedSingle, v16f32>, - PS, EVEX_V512, EVEX_CD8<32, CD8VF>; -defm VMOVAPDZ : avx512_load<0x28, VR512, VK8WM, f512mem, alignedloadv8f64, - "vmovapd", SSEPackedDouble, v8f64>, - avx512_store<0x29, VR512, VK8WM, f512mem, alignedstore512, - "vmovapd", SSEPackedDouble, v8f64>, - PD, EVEX_V512, VEX_W, - EVEX_CD8<64, CD8VF>; -defm VMOVUPSZ : avx512_load<0x10, VR512, VK16WM, f512mem, loadv16f32, - "vmovups", SSEPackedSingle, v16f32>, - avx512_store<0x11, VR512, VK16WM, f512mem, store, - "vmovups", SSEPackedSingle, v16f32>, - PS, EVEX_V512, EVEX_CD8<32, CD8VF>; -defm VMOVUPDZ : avx512_load<0x10, VR512, VK8WM, f512mem, loadv8f64, - "vmovupd", SSEPackedDouble, v8f64, 0>, - avx512_store<0x11, VR512, VK8WM, f512mem, store, - "vmovupd", SSEPackedDouble, v8f64>, - PD, EVEX_V512, VEX_W, - EVEX_CD8<64, CD8VF>; + +multiclass avx512_store_vl opc, string OpcodeStr, string st_pat, + string st_suff_512, string st_suff_256, + string st_suff_128, string elty, string elsz, + string vsz512, string vsz256, string vsz128, + Domain d, Predicate prd> { + let Predicates = [prd] in + defm Z : avx512_store(st_pat##st_suff_512), + !cast("v"##vsz512##elty##elsz), + !cast("VK"##vsz512##"WM"), VR512, + !cast(elty##"512mem"), d>, EVEX_V512; + + let Predicates = [prd, HasVLX] in { + defm Z256 : avx512_store(st_pat##st_suff_256), + !cast("v"##vsz256##elty##elsz), + !cast("VK"##vsz256##"WM"), VR256X, + !cast(elty##"256mem"), d>, EVEX_V256; + + defm Z128 : avx512_store(st_pat##st_suff_128), + !cast("v"##vsz128##elty##elsz), + !cast("VK"##vsz128##"WM"), VR128X, + !cast(elty##"128mem"), d>, EVEX_V128; + } +} + +defm VMOVAPS : avx512_load_vl<0x28, "vmovaps", "alignedload", "f", "32", + "16", "8", "4", SSEPackedSingle, HasAVX512>, + avx512_store_vl<0x29, "vmovaps", "alignedstore", + "512", "256", "", "f", "32", "16", "8", "4", + SSEPackedSingle, HasAVX512>, + PS, EVEX_CD8<32, CD8VF>; + +defm VMOVAPD : avx512_load_vl<0x28, "vmovapd", "alignedload", "f", "64", + "8", "4", "2", SSEPackedDouble, HasAVX512>, + avx512_store_vl<0x29, "vmovapd", "alignedstore", + "512", "256", "", "f", "64", "8", "4", "2", + SSEPackedDouble, HasAVX512>, + PD, VEX_W, EVEX_CD8<64, CD8VF>; + +defm VMOVUPS : avx512_load_vl<0x10, "vmovups", "load", "f", "32", + "16", "8", "4", SSEPackedSingle, HasAVX512>, + avx512_store_vl<0x11, "vmovups", "store", "", "", "", "f", "32", + "16", "8", "4", SSEPackedSingle, HasAVX512>, + PS, EVEX_CD8<32, CD8VF>; + +defm VMOVUPD : avx512_load_vl<0x10, "vmovupd", "load", "f", "64", + "8", "4", "2", SSEPackedDouble, HasAVX512, 0>, + avx512_store_vl<0x11, "vmovupd", "store", "", "", "", "f", "64", + "8", "4", "2", SSEPackedDouble, HasAVX512>, + PD, VEX_W, EVEX_CD8<64, CD8VF>; + def: Pat<(v8f64 (int_x86_avx512_mask_loadu_pd_512 addr:$ptr, - (bc_v8f64 (v16i32 immAllZerosV)), GR8:$mask)), + (bc_v8f64 (v16i32 immAllZerosV)), GR8:$mask)), (VMOVUPDZrmkz (v8i1 (COPY_TO_REGCLASS GR8:$mask, VK8WM)), addr:$ptr)>; def: Pat<(v16f32 (int_x86_avx512_mask_loadu_ps_512 addr:$ptr, @@ -1339,66 +2060,80 @@ def: Pat<(int_x86_avx512_mask_storeu_pd_512 addr:$ptr, (v8f64 VR512:$src), (VMOVUPDZmrk addr:$ptr, (v8i1 (COPY_TO_REGCLASS GR8:$mask, VK8WM)), VR512:$src)>; -defm VMOVDQA32: avx512_load<0x6F, VR512, VK16WM, i512mem, alignedloadv16i32, - "vmovdqa32", SSEPackedInt, v16i32>, - avx512_store<0x7F, VR512, VK16WM, i512mem, alignedstore512, - "vmovdqa32", SSEPackedInt, v16i32>, - PD, EVEX_V512, EVEX_CD8<32, CD8VF>; -defm VMOVDQA64: avx512_load<0x6F, VR512, VK8WM, i512mem, alignedloadv8i64, - "vmovdqa64", SSEPackedInt, v8i64>, - avx512_store<0x7F, VR512, VK8WM, i512mem, alignedstore512, - "vmovdqa64", SSEPackedInt, v8i64>, - PD, VEX_W, EVEX_V512, EVEX_CD8<64, CD8VF>; -defm VMOVDQU32: avx512_load<0x6F, VR512, VK16WM, i512mem, load, - "vmovdqu32", SSEPackedInt, v16i32>, - avx512_store<0x7F, VR512, VK16WM, i512mem, store, - "vmovdqu32", SSEPackedInt, v16i32>, - XS, EVEX_V512, EVEX_CD8<32, CD8VF>; -defm VMOVDQU64: avx512_load<0x6F, VR512, VK8WM, i512mem, load, - "vmovdqu64", SSEPackedInt, v8i64>, - avx512_store<0x7F, VR512, VK8WM, i512mem, store, - "vmovdqu64", SSEPackedInt, v8i64>, - XS, VEX_W, EVEX_V512, EVEX_CD8<64, CD8VF>; +defm VMOVDQA32 : avx512_load_vl<0x6F, "vmovdqa32", "alignedload", "i", "32", + "16", "8", "4", SSEPackedInt, HasAVX512>, + avx512_store_vl<0x7F, "vmovdqa32", "alignedstore", + "512", "256", "", "i", "32", "16", "8", "4", + SSEPackedInt, HasAVX512>, + PD, EVEX_CD8<32, CD8VF>; + +defm VMOVDQA64 : avx512_load_vl<0x6F, "vmovdqa64", "alignedload", "i", "64", + "8", "4", "2", SSEPackedInt, HasAVX512>, + avx512_store_vl<0x7F, "vmovdqa64", "alignedstore", + "512", "256", "", "i", "64", "8", "4", "2", + SSEPackedInt, HasAVX512>, + PD, VEX_W, EVEX_CD8<64, CD8VF>; + +defm VMOVDQU8 : avx512_load_vl<0x6F, "vmovdqu8", "load", "i", "8", + "64", "32", "16", SSEPackedInt, HasBWI>, + avx512_store_vl<0x7F, "vmovdqu8", "store", "", "", "", + "i", "8", "64", "32", "16", SSEPackedInt, + HasBWI>, XD, EVEX_CD8<8, CD8VF>; + +defm VMOVDQU16 : avx512_load_vl<0x6F, "vmovdqu16", "load", "i", "16", + "32", "16", "8", SSEPackedInt, HasBWI>, + avx512_store_vl<0x7F, "vmovdqu16", "store", "", "", "", + "i", "16", "32", "16", "8", SSEPackedInt, + HasBWI>, XD, VEX_W, EVEX_CD8<16, CD8VF>; + +defm VMOVDQU32 : avx512_load_vl<0x6F, "vmovdqu32", "load", "i", "32", + "16", "8", "4", SSEPackedInt, HasAVX512>, + avx512_store_vl<0x7F, "vmovdqu32", "store", "", "", "", + "i", "32", "16", "8", "4", SSEPackedInt, + HasAVX512>, XS, EVEX_CD8<32, CD8VF>; + +defm VMOVDQU64 : avx512_load_vl<0x6F, "vmovdqu64", "load", "i", "64", + "8", "4", "2", SSEPackedInt, HasAVX512>, + avx512_store_vl<0x7F, "vmovdqu64", "store", "", "", "", + "i", "64", "8", "4", "2", SSEPackedInt, + HasAVX512>, XS, VEX_W, EVEX_CD8<64, CD8VF>; def: Pat<(v16i32 (int_x86_avx512_mask_loadu_d_512 addr:$ptr, (v16i32 immAllZerosV), GR16:$mask)), - (VMOVDQU32rmkz (v16i1 (COPY_TO_REGCLASS GR16:$mask, VK16WM)), addr:$ptr)>; + (VMOVDQU32Zrmkz (v16i1 (COPY_TO_REGCLASS GR16:$mask, VK16WM)), addr:$ptr)>; def: Pat<(v8i64 (int_x86_avx512_mask_loadu_q_512 addr:$ptr, - (bc_v8i64 (v16i32 immAllZerosV)), GR8:$mask)), - (VMOVDQU64rmkz (v8i1 (COPY_TO_REGCLASS GR8:$mask, VK8WM)), addr:$ptr)>; + (bc_v8i64 (v16i32 immAllZerosV)), GR8:$mask)), + (VMOVDQU64Zrmkz (v8i1 (COPY_TO_REGCLASS GR8:$mask, VK8WM)), addr:$ptr)>; + +def: Pat<(int_x86_avx512_mask_storeu_d_512 addr:$ptr, (v16i32 VR512:$src), + GR16:$mask), + (VMOVDQU32Zmrk addr:$ptr, (v16i1 (COPY_TO_REGCLASS GR16:$mask, VK16WM)), + VR512:$src)>; +def: Pat<(int_x86_avx512_mask_storeu_q_512 addr:$ptr, (v8i64 VR512:$src), + GR8:$mask), + (VMOVDQU64Zmrk addr:$ptr, (v8i1 (COPY_TO_REGCLASS GR8:$mask, VK8WM)), + VR512:$src)>; let AddedComplexity = 20 in { def : Pat<(v8i64 (vselect VK8WM:$mask, (v8i64 VR512:$src), - (bc_v8i64 (v16i32 immAllZerosV)))), - (VMOVDQU64rrkz VK8WM:$mask, VR512:$src)>; + (bc_v8i64 (v16i32 immAllZerosV)))), + (VMOVDQU64Zrrkz VK8WM:$mask, VR512:$src)>; def : Pat<(v8i64 (vselect VK8WM:$mask, (bc_v8i64 (v16i32 immAllZerosV)), - (v8i64 VR512:$src))), - (VMOVDQU64rrkz (COPY_TO_REGCLASS (KNOTWrr (COPY_TO_REGCLASS VK8:$mask, VK16)), + (v8i64 VR512:$src))), + (VMOVDQU64Zrrkz (COPY_TO_REGCLASS (KNOTWrr (COPY_TO_REGCLASS VK8:$mask, VK16)), VK8), VR512:$src)>; def : Pat<(v16i32 (vselect VK16WM:$mask, (v16i32 VR512:$src), (v16i32 immAllZerosV))), - (VMOVDQU32rrkz VK16WM:$mask, VR512:$src)>; + (VMOVDQU32Zrrkz VK16WM:$mask, VR512:$src)>; def : Pat<(v16i32 (vselect VK16WM:$mask, (v16i32 immAllZerosV), - (v16i32 VR512:$src))), - (VMOVDQU32rrkz (KNOTWrr VK16WM:$mask), VR512:$src)>; - -def : Pat<(v16f32 (vselect VK16WM:$mask, (v16f32 VR512:$src1), - (v16f32 VR512:$src2))), - (VMOVUPSZrrk VR512:$src2, VK16WM:$mask, VR512:$src1)>; -def : Pat<(v8f64 (vselect VK8WM:$mask, (v8f64 VR512:$src1), - (v8f64 VR512:$src2))), - (VMOVUPDZrrk VR512:$src2, VK8WM:$mask, VR512:$src1)>; -def : Pat<(v16i32 (vselect VK16WM:$mask, (v16i32 VR512:$src1), - (v16i32 VR512:$src2))), - (VMOVDQU32rrk VR512:$src2, VK16WM:$mask, VR512:$src1)>; -def : Pat<(v8i64 (vselect VK8WM:$mask, (v8i64 VR512:$src1), - (v8i64 VR512:$src2))), - (VMOVDQU64rrk VR512:$src2, VK8WM:$mask, VR512:$src1)>; + (v16i32 VR512:$src))), + (VMOVDQU32Zrrkz (KNOTWrr VK16WM:$mask), VR512:$src)>; } + // Move Int Doubleword to Packed Double Int // def VMOVDI2PDIZrr : AVX512BI<0x6E, MRMSrcReg, (outs VR128X:$dst), (ins GR32:$src), @@ -1524,10 +2259,16 @@ multiclass avx512_move_scalar , EVEX, VEX_LIG; + 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}}, $src}"), + [], IIC_SSE_MOV_S_MR>, + EVEX, VEX_LIG, EVEX_K; + } // mayStore } //hasSideEffects = 0 } @@ -1547,6 +2288,10 @@ def : Pat<(f64 (X86select VK1WM:$mask, (f64 FR64X:$src1), (f64 FR64X:$src2))), (COPY_TO_REGCLASS (VMOVSDZrrk (COPY_TO_REGCLASS FR64X:$src2, VR128X), VK1WM:$mask, (f64 (IMPLICIT_DEF)), FR64X:$src1), FR64X)>; +def : Pat<(int_x86_avx512_mask_store_ss addr:$dst, VR128X:$src, GR8:$mask), + (VMOVSSZmrk addr:$dst, (i1 (COPY_TO_REGCLASS GR8:$mask, VK1WM)), + (COPY_TO_REGCLASS VR128X:$src, FR32X))>; + // For the disassembler let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0 in { def VMOVSSZrr_REV : SI<0x11, MRMDestReg, (outs VR128X:$dst), @@ -1762,99 +2507,204 @@ def : Pat<(v16i32 (X86Vinsert undef, GR32:$src2, (iPTR 0))), def : Pat<(v8i64 (X86Vinsert undef, GR64:$src2, (iPTR 0))), (SUBREG_TO_REG (i32 0), (VMOV64toPQIZrr GR64:$src2), sub_xmm)>; +//===----------------------------------------------------------------------===// +// AVX-512 - Non-temporals +//===----------------------------------------------------------------------===// +let SchedRW = [WriteLoad] in { + def VMOVNTDQAZrm : AVX512PI<0x2A, MRMSrcMem, (outs VR512:$dst), + (ins i512mem:$src), "vmovntdqa\t{$src, $dst|$dst, $src}", + [(set VR512:$dst, (int_x86_avx512_movntdqa addr:$src))], + SSEPackedInt>, EVEX, T8PD, EVEX_V512, + EVEX_CD8<64, CD8VF>; + + let Predicates = [HasAVX512, HasVLX] in { + def VMOVNTDQAZ256rm : AVX512PI<0x2A, MRMSrcMem, (outs VR256X:$dst), + (ins i256mem:$src), + "vmovntdqa\t{$src, $dst|$dst, $src}", [], + SSEPackedInt>, EVEX, T8PD, EVEX_V256, + EVEX_CD8<64, CD8VF>; + + def VMOVNTDQAZ128rm : AVX512PI<0x2A, MRMSrcMem, (outs VR128X:$dst), + (ins i128mem:$src), + "vmovntdqa\t{$src, $dst|$dst, $src}", [], + SSEPackedInt>, EVEX, T8PD, EVEX_V128, + EVEX_CD8<64, CD8VF>; + } +} + +multiclass avx512_movnt opc, string OpcodeStr, PatFrag st_frag, + ValueType OpVT, RegisterClass RC, X86MemOperand memop, + Domain d, InstrItinClass itin = IIC_SSE_MOVNT> { + let SchedRW = [WriteStore], mayStore = 1, + AddedComplexity = 400 in + def mr : AVX512PI, EVEX; +} + +multiclass avx512_movnt_vl opc, string OpcodeStr, PatFrag st_frag, + string elty, string elsz, string vsz512, + string vsz256, string vsz128, Domain d, + Predicate prd, InstrItinClass itin = IIC_SSE_MOVNT> { + let Predicates = [prd] in + defm Z : avx512_movnt("v"##vsz512##elty##elsz), VR512, + !cast(elty##"512mem"), d, itin>, + EVEX_V512; + + let Predicates = [prd, HasVLX] in { + defm Z256 : avx512_movnt("v"##vsz256##elty##elsz), VR256X, + !cast(elty##"256mem"), d, itin>, + EVEX_V256; + + defm Z128 : avx512_movnt("v"##vsz128##elty##elsz), VR128X, + !cast(elty##"128mem"), d, itin>, + EVEX_V128; + } +} + +defm VMOVNTDQ : avx512_movnt_vl<0xE7, "vmovntdq", alignednontemporalstore, + "i", "64", "8", "4", "2", SSEPackedInt, + HasAVX512>, PD, EVEX_CD8<64, CD8VF>; + +defm VMOVNTPD : avx512_movnt_vl<0x2B, "vmovntpd", alignednontemporalstore, + "f", "64", "8", "4", "2", SSEPackedDouble, + HasAVX512>, PD, VEX_W, EVEX_CD8<64, CD8VF>; + +defm VMOVNTPS : avx512_movnt_vl<0x2B, "vmovntps", alignednontemporalstore, + "f", "32", "16", "8", "4", SSEPackedSingle, + HasAVX512>, PS, EVEX_CD8<32, CD8VF>; + //===----------------------------------------------------------------------===// // AVX-512 - Integer arithmetic // multiclass avx512_binop_rm opc, string OpcodeStr, SDNode OpNode, - ValueType OpVT, RegisterClass KRC, - RegisterClass RC, PatFrag memop_frag, - X86MemOperand x86memop, PatFrag scalar_mfrag, - X86MemOperand x86scalar_mop, string BrdcstStr, - OpndItins itins, bit IsCommutable = 0> { - let isCommutable = IsCommutable in - def rr : AVX512BI, EVEX_4V; - let AddedComplexity = 30 in { - let Constraints = "$src0 = $dst" in - def rrk : AVX512BI, EVEX_4V, EVEX_K; - def rrkz : AVX512BI, EVEX_4V, EVEX_KZ; + X86VectorVTInfo _, OpndItins itins, + bit IsCommutable = 0> { + defm rr : AVX512_maskable, + AVX512BIBase, EVEX_4V; + + let mayLoad = 1 in + defm rm : AVX512_maskable, + AVX512BIBase, EVEX_4V; +} + +multiclass avx512_binop_rmb opc, string OpcodeStr, SDNode OpNode, + X86VectorVTInfo _, OpndItins itins, + bit IsCommutable = 0> : + avx512_binop_rm { + let mayLoad = 1 in + defm rmb : AVX512_maskable, + AVX512BIBase, EVEX_4V, EVEX_B; +} + +multiclass avx512_binop_rm_vl opc, string OpcodeStr, SDNode OpNode, + AVX512VLVectorVTInfo VTInfo, OpndItins itins, + Predicate prd, bit IsCommutable = 0> { + let Predicates = [prd] in + defm Z : avx512_binop_rm, EVEX_V512; + + let Predicates = [prd, HasVLX] in { + defm Z256 : avx512_binop_rm, EVEX_V256; + defm Z128 : avx512_binop_rm, EVEX_V128; } +} - let mayLoad = 1 in { - def rm : AVX512BI, EVEX_4V; - let AddedComplexity = 30 in { - let Constraints = "$src0 = $dst" in - def rmk : AVX512BI, EVEX_4V, EVEX_K; - def rmkz : AVX512BI, EVEX_4V, EVEX_KZ; - } - def rmb : AVX512BI, EVEX_4V, EVEX_B; - let AddedComplexity = 30 in { - let Constraints = "$src0 = $dst" in - def rmbk : AVX512BI, EVEX_4V, EVEX_B, EVEX_K; - def rmbkz : AVX512BI, EVEX_4V, EVEX_B, EVEX_KZ; - } +multiclass avx512_binop_rmb_vl opc, string OpcodeStr, SDNode OpNode, + AVX512VLVectorVTInfo VTInfo, OpndItins itins, + Predicate prd, bit IsCommutable = 0> { + let Predicates = [prd] in + defm Z : avx512_binop_rmb, EVEX_V512; + + let Predicates = [prd, HasVLX] in { + defm Z256 : avx512_binop_rmb, EVEX_V256; + defm Z128 : avx512_binop_rmb, EVEX_V128; } } +multiclass avx512_binop_rm_vl_q opc, string OpcodeStr, SDNode OpNode, + OpndItins itins, Predicate prd, + bit IsCommutable = 0> { + defm NAME : avx512_binop_rmb_vl, + VEX_W, EVEX_CD8<64, CD8VF>; +} + +multiclass avx512_binop_rm_vl_d opc, string OpcodeStr, SDNode OpNode, + OpndItins itins, Predicate prd, + bit IsCommutable = 0> { + defm NAME : avx512_binop_rmb_vl, EVEX_CD8<32, CD8VF>; +} + +multiclass avx512_binop_rm_vl_w opc, string OpcodeStr, SDNode OpNode, + OpndItins itins, Predicate prd, + bit IsCommutable = 0> { + defm NAME : avx512_binop_rm_vl, EVEX_CD8<16, CD8VF>; +} + +multiclass avx512_binop_rm_vl_b opc, string OpcodeStr, SDNode OpNode, + OpndItins itins, Predicate prd, + bit IsCommutable = 0> { + defm NAME : avx512_binop_rm_vl, EVEX_CD8<8, CD8VF>; +} + +multiclass avx512_binop_rm_vl_dq opc_d, bits<8> opc_q, string OpcodeStr, + SDNode OpNode, OpndItins itins, Predicate prd, + bit IsCommutable = 0> { + defm Q : avx512_binop_rm_vl_q; + + defm D : avx512_binop_rm_vl_d; +} + +multiclass avx512_binop_rm_vl_bw opc_b, bits<8> opc_w, string OpcodeStr, + SDNode OpNode, OpndItins itins, Predicate prd, + bit IsCommutable = 0> { + defm W : avx512_binop_rm_vl_w; + + defm B : avx512_binop_rm_vl_b; +} + +multiclass avx512_binop_rm_vl_all opc_b, bits<8> opc_w, + bits<8> opc_d, bits<8> opc_q, + string OpcodeStr, SDNode OpNode, + OpndItins itins, bit IsCommutable = 0> { + defm NAME : avx512_binop_rm_vl_dq, + avx512_binop_rm_vl_bw; +} + multiclass avx512_binop_rm2 opc, string OpcodeStr, ValueType DstVT, ValueType SrcVT, RegisterClass KRC, RegisterClass RC, PatFrag memop_frag, X86MemOperand x86memop, @@ -1912,25 +2762,16 @@ multiclass avx512_binop_rm2 opc, string OpcodeStr, ValueType DstVT, } } -defm VPADDDZ : avx512_binop_rm<0xFE, "vpaddd", add, v16i32, VK16WM, VR512, - memopv16i32, i512mem, loadi32, i32mem, "{1to16}", - SSE_INTALU_ITINS_P, 1>, EVEX_V512, EVEX_CD8<32, CD8VF>; - -defm VPSUBDZ : avx512_binop_rm<0xFA, "vpsubd", sub, v16i32, VK16WM, VR512, - memopv16i32, i512mem, loadi32, i32mem, "{1to16}", - SSE_INTALU_ITINS_P, 0>, EVEX_V512, EVEX_CD8<32, CD8VF>; - -defm VPMULLDZ : avx512_binop_rm<0x40, "vpmulld", mul, v16i32, VK16WM, VR512, - memopv16i32, i512mem, loadi32, i32mem, "{1to16}", - SSE_INTALU_ITINS_P, 1>, T8PD, EVEX_V512, EVEX_CD8<32, CD8VF>; - -defm VPADDQZ : avx512_binop_rm<0xD4, "vpaddq", add, v8i64, VK8WM, VR512, - memopv8i64, i512mem, loadi64, i64mem, "{1to8}", - SSE_INTALU_ITINS_P, 1>, EVEX_CD8<64, CD8VF>, EVEX_V512, VEX_W; - -defm VPSUBQZ : avx512_binop_rm<0xFB, "vpsubq", sub, v8i64, VK8WM, VR512, - memopv8i64, i512mem, loadi64, i64mem, "{1to8}", - SSE_INTALU_ITINS_P, 0>, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>; +defm VPADD : avx512_binop_rm_vl_all<0xFC, 0xFD, 0xFE, 0xD4, "vpadd", add, + SSE_INTALU_ITINS_P, 1>; +defm VPSUB : avx512_binop_rm_vl_all<0xF8, 0xF9, 0xFA, 0xFB, "vpsub", sub, + SSE_INTALU_ITINS_P, 0>; +defm VPMULLD : avx512_binop_rm_vl_d<0x40, "vpmull", mul, + SSE_INTALU_ITINS_P, HasAVX512, 1>, T8PD; +defm VPMULLW : avx512_binop_rm_vl_w<0xD5, "vpmull", mul, + SSE_INTALU_ITINS_P, HasBWI, 1>; +defm VPMULLQ : avx512_binop_rm_vl_q<0x40, "vpmull", mul, + SSE_INTALU_ITINS_P, HasDQI, 1>, T8PD; defm VPMULDQZ : avx512_binop_rm2<0x28, "vpmuldq", v8i64, v16i32, VK8WM, VR512, memopv8i64, i512mem, loadi64, i64mem, "{1to8}", @@ -1951,41 +2792,33 @@ def : Pat<(v8i64 (int_x86_avx512_mask_pmul_dq_512 (v16i32 VR512:$src1), (v16i32 VR512:$src2), (bc_v8i64 (v16i32 immAllZerosV)), (i8 -1))), (VPMULDQZrr VR512:$src1, VR512:$src2)>; -defm VPMAXUDZ : avx512_binop_rm<0x3F, "vpmaxud", X86umax, v16i32, VK16WM, VR512, - memopv16i32, i512mem, loadi32, i32mem, "{1to16}", - SSE_INTALU_ITINS_P, 1>, - T8PD, EVEX_V512, EVEX_CD8<32, CD8VF>; -defm VPMAXUQZ : avx512_binop_rm<0x3F, "vpmaxuq", X86umax, v8i64, VK8WM, VR512, - memopv8i64, i512mem, loadi64, i64mem, "{1to8}", - SSE_INTALU_ITINS_P, 0>, - T8PD, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>; - -defm VPMAXSDZ : avx512_binop_rm<0x3D, "vpmaxsd", X86smax, v16i32, VK16WM, VR512, - memopv16i32, i512mem, loadi32, i32mem, "{1to16}", - SSE_INTALU_ITINS_P, 1>, - T8PD, EVEX_V512, EVEX_CD8<32, CD8VF>; -defm VPMAXSQZ : avx512_binop_rm<0x3D, "vpmaxsq", X86smax, v8i64, VK8WM, VR512, - memopv8i64, i512mem, loadi64, i64mem, "{1to8}", - SSE_INTALU_ITINS_P, 0>, - T8PD, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>; - -defm VPMINUDZ : avx512_binop_rm<0x3B, "vpminud", X86umin, v16i32, VK16WM, VR512, - memopv16i32, i512mem, loadi32, i32mem, "{1to16}", - SSE_INTALU_ITINS_P, 1>, - T8PD, EVEX_V512, EVEX_CD8<32, CD8VF>; -defm VPMINUQZ : avx512_binop_rm<0x3B, "vpminuq", X86umin, v8i64, VK8WM, VR512, - memopv8i64, i512mem, loadi64, i64mem, "{1to8}", - SSE_INTALU_ITINS_P, 0>, - T8PD, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>; - -defm VPMINSDZ : avx512_binop_rm<0x39, "vpminsd", X86smin, v16i32, VK16WM, VR512, - memopv16i32, i512mem, loadi32, i32mem, "{1to16}", - SSE_INTALU_ITINS_P, 1>, - T8PD, EVEX_V512, EVEX_CD8<32, CD8VF>; -defm VPMINSQZ : avx512_binop_rm<0x39, "vpminsq", X86smin, v8i64, VK8WM, VR512, - memopv8i64, i512mem, loadi64, i64mem, "{1to8}", - SSE_INTALU_ITINS_P, 0>, - T8PD, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>; +defm VPMAXSB : avx512_binop_rm_vl_b<0x3C, "vpmaxs", X86smax, + SSE_INTALU_ITINS_P, HasBWI, 1>, T8PD; +defm VPMAXSW : avx512_binop_rm_vl_w<0xEE, "vpmaxs", X86smax, + SSE_INTALU_ITINS_P, HasBWI, 1>; +defm VPMAXS : avx512_binop_rm_vl_dq<0x3D, 0x3D, "vpmaxs", X86smax, + SSE_INTALU_ITINS_P, HasAVX512, 1>, T8PD; + +defm VPMAXUB : avx512_binop_rm_vl_b<0xDE, "vpmaxu", X86umax, + SSE_INTALU_ITINS_P, HasBWI, 1>; +defm VPMAXUW : avx512_binop_rm_vl_w<0x3E, "vpmaxu", X86umax, + SSE_INTALU_ITINS_P, HasBWI, 1>, T8PD; +defm VPMAXU : avx512_binop_rm_vl_dq<0x3F, 0x3F, "vpmaxu", X86umax, + SSE_INTALU_ITINS_P, HasAVX512, 1>, T8PD; + +defm VPMINSB : avx512_binop_rm_vl_b<0x38, "vpmins", X86smin, + SSE_INTALU_ITINS_P, HasBWI, 1>, T8PD; +defm VPMINSW : avx512_binop_rm_vl_w<0xEA, "vpmins", X86smin, + SSE_INTALU_ITINS_P, HasBWI, 1>; +defm VPMINS : avx512_binop_rm_vl_dq<0x39, 0x39, "vpmins", X86smin, + SSE_INTALU_ITINS_P, HasAVX512, 1>, T8PD; + +defm VPMINUB : avx512_binop_rm_vl_b<0xDA, "vpminu", X86umin, + SSE_INTALU_ITINS_P, HasBWI, 1>; +defm VPMINUW : avx512_binop_rm_vl_w<0x3A, "vpminu", X86umin, + SSE_INTALU_ITINS_P, HasBWI, 1>, T8PD; +defm VPMINU : avx512_binop_rm_vl_dq<0x3B, 0x3B, "vpminu", X86umin, + SSE_INTALU_ITINS_P, HasAVX512, 1>, T8PD; def : Pat <(v16i32 (int_x86_avx512_mask_pmaxs_d_512 (v16i32 VR512:$src1), (v16i32 VR512:$src2), (v16i32 immAllZerosV), (i16 -1))), @@ -2098,48 +2931,18 @@ multiclass avx512_pshuf_imm opc, string OpcodeStr, RegisterClass RC, defm VPSHUFDZ : avx512_pshuf_imm<0x70, "vpshufd", VR512, X86PShufd, memopv16i32, i512mem, v16i32>, PD, EVEX_V512, EVEX_CD8<32, CD8VF>; -let ExeDomain = SSEPackedSingle in -defm VPERMILPSZ : avx512_pshuf_imm<0x04, "vpermilps", VR512, X86VPermilp, - memopv16f32, i512mem, v16f32>, TAPD, EVEX_V512, - EVEX_CD8<32, CD8VF>; -let ExeDomain = SSEPackedDouble in -defm VPERMILPDZ : avx512_pshuf_imm<0x05, "vpermilpd", VR512, X86VPermilp, - memopv8f64, i512mem, v8f64>, TAPD, EVEX_V512, - VEX_W, EVEX_CD8<32, CD8VF>; - -def : Pat<(v16i32 (X86VPermilp VR512:$src1, (i8 imm:$imm))), - (VPERMILPSZri VR512:$src1, imm:$imm)>; -def : Pat<(v8i64 (X86VPermilp VR512:$src1, (i8 imm:$imm))), - (VPERMILPDZri VR512:$src1, imm:$imm)>; - //===----------------------------------------------------------------------===// // AVX-512 Logical Instructions //===----------------------------------------------------------------------===// -defm VPANDDZ : avx512_binop_rm<0xDB, "vpandd", and, v16i32, VK16WM, VR512, memopv16i32, - i512mem, loadi32, i32mem, "{1to16}", SSE_BIT_ITINS_P, 1>, - EVEX_V512, EVEX_CD8<32, CD8VF>; -defm VPANDQZ : avx512_binop_rm<0xDB, "vpandq", and, v8i64, VK8WM, VR512, memopv8i64, - i512mem, loadi64, i64mem, "{1to8}", SSE_BIT_ITINS_P, 1>, - EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>; -defm VPORDZ : avx512_binop_rm<0xEB, "vpord", or, v16i32, VK16WM, VR512, memopv16i32, - i512mem, loadi32, i32mem, "{1to16}", SSE_BIT_ITINS_P, 1>, - EVEX_V512, EVEX_CD8<32, CD8VF>; -defm VPORQZ : avx512_binop_rm<0xEB, "vporq", or, v8i64, VK8WM, VR512, memopv8i64, - i512mem, loadi64, i64mem, "{1to8}", SSE_BIT_ITINS_P, 1>, - EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>; -defm VPXORDZ : avx512_binop_rm<0xEF, "vpxord", xor, v16i32, VK16WM, VR512, memopv16i32, - i512mem, loadi32, i32mem, "{1to16}", SSE_BIT_ITINS_P, 1>, - EVEX_V512, EVEX_CD8<32, CD8VF>; -defm VPXORQZ : avx512_binop_rm<0xEF, "vpxorq", xor, v8i64, VK8WM, VR512, memopv8i64, - i512mem, loadi64, i64mem, "{1to8}", SSE_BIT_ITINS_P, 1>, - EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>; -defm VPANDNDZ : avx512_binop_rm<0xDF, "vpandnd", X86andnp, v16i32, VK16WM, VR512, - memopv16i32, i512mem, loadi32, i32mem, "{1to16}", - SSE_BIT_ITINS_P, 0>, EVEX_V512, EVEX_CD8<32, CD8VF>; -defm VPANDNQZ : avx512_binop_rm<0xDF, "vpandnq", X86andnp, v8i64, VK8WM, VR512, - memopv8i64, i512mem, loadi64, i64mem, "{1to8}", - SSE_BIT_ITINS_P, 0>, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>; +defm VPAND : avx512_binop_rm_vl_dq<0xDB, 0xDB, "vpand", and, + SSE_INTALU_ITINS_P, HasAVX512, 1>; +defm VPOR : avx512_binop_rm_vl_dq<0xEB, 0xEB, "vpor", or, + SSE_INTALU_ITINS_P, HasAVX512, 1>; +defm VPXOR : avx512_binop_rm_vl_dq<0xEF, 0xEF, "vpxor", xor, + SSE_INTALU_ITINS_P, HasAVX512, 1>; +defm VPANDN : avx512_binop_rm_vl_dq<0xDF, 0xDF, "vpandn", X86andnp, + SSE_INTALU_ITINS_P, HasAVX512, 1>; //===----------------------------------------------------------------------===// // AVX-512 FP arithmetic @@ -2556,155 +3359,133 @@ let Predicates = [HasAVX512] in { //===----------------------------------------------------------------------===// // FMA - Fused Multiply Operations // + let Constraints = "$src1 = $dst" in { -multiclass avx512_fma3p_rm opc, string OpcodeStr, - RegisterClass RC, X86MemOperand x86memop, - PatFrag mem_frag, X86MemOperand x86scalar_mop, PatFrag scalar_mfrag, - string BrdcstStr, SDNode OpNode, ValueType OpVT> { - def r: AVX512FMA3; +// Omitting the parameter OpNode (= null_frag) disables ISel pattern matching. +multiclass avx512_fma3p_rm opc, string OpcodeStr, X86VectorVTInfo _, + SDPatternOperator OpNode = null_frag> { + defm r: AVX512_maskable_3src, + AVX512FMA3Base; let mayLoad = 1 in - def m: AVX512FMA3; - def mb: AVX512FMA3, EVEX_B; + [(set _.RC:$dst, (_.VT (OpNode _.RC:$src1, _.RC:$src2, + (_.MemOpFrag addr:$src3))))]>; + def mb: AVX512FMA3, EVEX_B; } } // Constraints = "$src1 = $dst" +multiclass avx512_fma3p_forms opc213, bits<8> opc231, + string OpcodeStr, X86VectorVTInfo VTI, + SDPatternOperator OpNode> { + defm v213 : avx512_fma3p_rm, + EVEX_V512, EVEX_CD8; + + defm v231 : avx512_fma3p_rm, + EVEX_V512, EVEX_CD8; +} + let ExeDomain = SSEPackedSingle in { - defm VFMADD213PSZ : avx512_fma3p_rm<0xA8, "vfmadd213ps", VR512, f512mem, - memopv16f32, f32mem, loadf32, "{1to16}", - X86Fmadd, v16f32>, EVEX_V512, - EVEX_CD8<32, CD8VF>; - defm VFMSUB213PSZ : avx512_fma3p_rm<0xAA, "vfmsub213ps", VR512, f512mem, - memopv16f32, f32mem, loadf32, "{1to16}", - X86Fmsub, v16f32>, EVEX_V512, - EVEX_CD8<32, CD8VF>; - defm VFMADDSUB213PSZ : avx512_fma3p_rm<0xA6, "vfmaddsub213ps", VR512, f512mem, - memopv16f32, f32mem, loadf32, "{1to16}", - X86Fmaddsub, v16f32>, - EVEX_V512, EVEX_CD8<32, CD8VF>; - defm VFMSUBADD213PSZ : avx512_fma3p_rm<0xA7, "vfmsubadd213ps", VR512, f512mem, - memopv16f32, f32mem, loadf32, "{1to16}", - X86Fmsubadd, v16f32>, - EVEX_V512, EVEX_CD8<32, CD8VF>; - defm VFNMADD213PSZ : avx512_fma3p_rm<0xAC, "vfnmadd213ps", VR512, f512mem, - memopv16f32, f32mem, loadf32, "{1to16}", - X86Fnmadd, v16f32>, EVEX_V512, - EVEX_CD8<32, CD8VF>; - defm VFNMSUB213PSZ : avx512_fma3p_rm<0xAE, "vfnmsub213ps", VR512, f512mem, - memopv16f32, f32mem, loadf32, "{1to16}", - X86Fnmsub, v16f32>, EVEX_V512, - EVEX_CD8<32, CD8VF>; + defm VFMADDPSZ : avx512_fma3p_forms<0xA8, 0xB8, "vfmadd", + v16f32_info, X86Fmadd>; + defm VFMSUBPSZ : avx512_fma3p_forms<0xAA, 0xBA, "vfmsub", + v16f32_info, X86Fmsub>; + defm VFMADDSUBPSZ : avx512_fma3p_forms<0xA6, 0xB6, "vfmaddsub", + v16f32_info, X86Fmaddsub>; + defm VFMSUBADDPSZ : avx512_fma3p_forms<0xA7, 0xB7, "vfmsubadd", + v16f32_info, X86Fmsubadd>; + defm VFNMADDPSZ : avx512_fma3p_forms<0xAC, 0xBC, "vfnmadd", + v16f32_info, X86Fnmadd>; + defm VFNMSUBPSZ : avx512_fma3p_forms<0xAE, 0xBE, "vfnmsub", + v16f32_info, X86Fnmsub>; } let ExeDomain = SSEPackedDouble in { - defm VFMADD213PDZ : avx512_fma3p_rm<0xA8, "vfmadd213pd", VR512, f512mem, - memopv8f64, f64mem, loadf64, "{1to8}", - X86Fmadd, v8f64>, EVEX_V512, - VEX_W, EVEX_CD8<64, CD8VF>; - defm VFMSUB213PDZ : avx512_fma3p_rm<0xAA, "vfmsub213pd", VR512, f512mem, - memopv8f64, f64mem, loadf64, "{1to8}", - X86Fmsub, v8f64>, EVEX_V512, VEX_W, - EVEX_CD8<64, CD8VF>; - defm VFMADDSUB213PDZ : avx512_fma3p_rm<0xA6, "vfmaddsub213pd", VR512, f512mem, - memopv8f64, f64mem, loadf64, "{1to8}", - X86Fmaddsub, v8f64>, EVEX_V512, VEX_W, - EVEX_CD8<64, CD8VF>; - defm VFMSUBADD213PDZ : avx512_fma3p_rm<0xA7, "vfmsubadd213pd", VR512, f512mem, - memopv8f64, f64mem, loadf64, "{1to8}", - X86Fmsubadd, v8f64>, EVEX_V512, VEX_W, - EVEX_CD8<64, CD8VF>; - defm VFNMADD213PDZ : avx512_fma3p_rm<0xAC, "vfnmadd213pd", VR512, f512mem, - memopv8f64, f64mem, loadf64, "{1to8}", - X86Fnmadd, v8f64>, EVEX_V512, VEX_W, - EVEX_CD8<64, CD8VF>; - defm VFNMSUB213PDZ : avx512_fma3p_rm<0xAE, "vfnmsub213pd", VR512, f512mem, - memopv8f64, f64mem, loadf64, "{1to8}", - X86Fnmsub, v8f64>, EVEX_V512, VEX_W, - EVEX_CD8<64, CD8VF>; + defm VFMADDPDZ : avx512_fma3p_forms<0xA8, 0xB8, "vfmadd", + v8f64_info, X86Fmadd>, VEX_W; + defm VFMSUBPDZ : avx512_fma3p_forms<0xAA, 0xBA, "vfmsub", + v8f64_info, X86Fmsub>, VEX_W; + defm VFMADDSUBPDZ : avx512_fma3p_forms<0xA6, 0xB6, "vfmaddsub", + v8f64_info, X86Fmaddsub>, VEX_W; + defm VFMSUBADDPDZ : avx512_fma3p_forms<0xA7, 0xB7, "vfmsubadd", + v8f64_info, X86Fmsubadd>, VEX_W; + defm VFNMADDPDZ : avx512_fma3p_forms<0xAC, 0xBC, "vfnmadd", + v8f64_info, X86Fnmadd>, VEX_W; + defm VFNMSUBPDZ : avx512_fma3p_forms<0xAE, 0xBE, "vfnmsub", + v8f64_info, X86Fnmsub>, VEX_W; } let Constraints = "$src1 = $dst" in { -multiclass avx512_fma3p_m132 opc, string OpcodeStr, - RegisterClass RC, X86MemOperand x86memop, - PatFrag mem_frag, X86MemOperand x86scalar_mop, PatFrag scalar_mfrag, - string BrdcstStr, SDNode OpNode, ValueType OpVT> { +multiclass avx512_fma3p_m132 opc, string OpcodeStr, SDNode OpNode, + X86VectorVTInfo _> { let mayLoad = 1 in - def m: AVX512FMA3; - def mb: AVX512FMA3, EVEX_B; + [(set _.RC:$dst, (_.VT (OpNode _.RC:$src1, (_.MemOpFrag addr:$src2), + _.RC:$src3)))]>; + def mb: AVX512FMA3, EVEX_B; } } // Constraints = "$src1 = $dst" let ExeDomain = SSEPackedSingle in { - defm VFMADD132PSZ : avx512_fma3p_m132<0x98, "vfmadd132ps", VR512, f512mem, - memopv16f32, f32mem, loadf32, "{1to16}", - X86Fmadd, v16f32>, EVEX_V512, - EVEX_CD8<32, CD8VF>; - defm VFMSUB132PSZ : avx512_fma3p_m132<0x9A, "vfmsub132ps", VR512, f512mem, - memopv16f32, f32mem, loadf32, "{1to16}", - X86Fmsub, v16f32>, EVEX_V512, - EVEX_CD8<32, CD8VF>; - defm VFMADDSUB132PSZ : avx512_fma3p_m132<0x96, "vfmaddsub132ps", VR512, f512mem, - memopv16f32, f32mem, loadf32, "{1to16}", - X86Fmaddsub, v16f32>, - EVEX_V512, EVEX_CD8<32, CD8VF>; - defm VFMSUBADD132PSZ : avx512_fma3p_m132<0x97, "vfmsubadd132ps", VR512, f512mem, - memopv16f32, f32mem, loadf32, "{1to16}", - X86Fmsubadd, v16f32>, - EVEX_V512, EVEX_CD8<32, CD8VF>; - defm VFNMADD132PSZ : avx512_fma3p_m132<0x9C, "vfnmadd132ps", VR512, f512mem, - memopv16f32, f32mem, loadf32, "{1to16}", - X86Fnmadd, v16f32>, EVEX_V512, - EVEX_CD8<32, CD8VF>; - defm VFNMSUB132PSZ : avx512_fma3p_m132<0x9E, "vfnmsub132ps", VR512, f512mem, - memopv16f32, f32mem, loadf32, "{1to16}", - X86Fnmsub, v16f32>, EVEX_V512, - EVEX_CD8<32, CD8VF>; + defm VFMADD132PSZ : avx512_fma3p_m132<0x98, "vfmadd132ps", X86Fmadd, + v16f32_info>, + EVEX_V512, EVEX_CD8<32, CD8VF>; + defm VFMSUB132PSZ : avx512_fma3p_m132<0x9A, "vfmsub132ps", X86Fmsub, + v16f32_info>, + EVEX_V512, EVEX_CD8<32, CD8VF>; + defm VFMADDSUB132PSZ : avx512_fma3p_m132<0x96, "vfmaddsub132ps", X86Fmaddsub, + v16f32_info>, + EVEX_V512, EVEX_CD8<32, CD8VF>; + defm VFMSUBADD132PSZ : avx512_fma3p_m132<0x97, "vfmsubadd132ps", X86Fmsubadd, + v16f32_info>, + EVEX_V512, EVEX_CD8<32, CD8VF>; + defm VFNMADD132PSZ : avx512_fma3p_m132<0x9C, "vfnmadd132ps", X86Fnmadd, + v16f32_info>, + EVEX_V512, EVEX_CD8<32, CD8VF>; + defm VFNMSUB132PSZ : avx512_fma3p_m132<0x9E, "vfnmsub132ps", X86Fnmsub, + v16f32_info>, + EVEX_V512, EVEX_CD8<32, CD8VF>; } let ExeDomain = SSEPackedDouble in { - defm VFMADD132PDZ : avx512_fma3p_m132<0x98, "vfmadd132pd", VR512, f512mem, - memopv8f64, f64mem, loadf64, "{1to8}", - X86Fmadd, v8f64>, EVEX_V512, - VEX_W, EVEX_CD8<64, CD8VF>; - defm VFMSUB132PDZ : avx512_fma3p_m132<0x9A, "vfmsub132pd", VR512, f512mem, - memopv8f64, f64mem, loadf64, "{1to8}", - X86Fmsub, v8f64>, EVEX_V512, VEX_W, - EVEX_CD8<64, CD8VF>; - defm VFMADDSUB132PDZ : avx512_fma3p_m132<0x96, "vfmaddsub132pd", VR512, f512mem, - memopv8f64, f64mem, loadf64, "{1to8}", - X86Fmaddsub, v8f64>, EVEX_V512, VEX_W, - EVEX_CD8<64, CD8VF>; - defm VFMSUBADD132PDZ : avx512_fma3p_m132<0x97, "vfmsubadd132pd", VR512, f512mem, - memopv8f64, f64mem, loadf64, "{1to8}", - X86Fmsubadd, v8f64>, EVEX_V512, VEX_W, - EVEX_CD8<64, CD8VF>; - defm VFNMADD132PDZ : avx512_fma3p_m132<0x9C, "vfnmadd132pd", VR512, f512mem, - memopv8f64, f64mem, loadf64, "{1to8}", - X86Fnmadd, v8f64>, EVEX_V512, VEX_W, - EVEX_CD8<64, CD8VF>; - defm VFNMSUB132PDZ : avx512_fma3p_m132<0x9E, "vfnmsub132pd", VR512, f512mem, - memopv8f64, f64mem, loadf64, "{1to8}", - X86Fnmsub, v8f64>, EVEX_V512, VEX_W, - EVEX_CD8<64, CD8VF>; + defm VFMADD132PDZ : avx512_fma3p_m132<0x98, "vfmadd132pd", X86Fmadd, + v8f64_info>, + EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>; + defm VFMSUB132PDZ : avx512_fma3p_m132<0x9A, "vfmsub132pd", X86Fmsub, + v8f64_info>, + EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>; + defm VFMADDSUB132PDZ : avx512_fma3p_m132<0x96, "vfmaddsub132pd", X86Fmaddsub, + v8f64_info>, + EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>; + defm VFMSUBADD132PDZ : avx512_fma3p_m132<0x97, "vfmsubadd132pd", X86Fmsubadd, + v8f64_info>, + EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>; + defm VFNMADD132PDZ : avx512_fma3p_m132<0x9C, "vfnmadd132pd", X86Fnmadd, + v8f64_info>, + EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>; + defm VFNMSUB132PDZ : avx512_fma3p_m132<0x9E, "vfnmsub132pd", X86Fnmsub, + v8f64_info>, + EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>; } // Scalar FMA @@ -3136,6 +3917,10 @@ def : Pat<(v4f32 (uint_to_fp (v4i32 VR128X:$src1))), (EXTRACT_SUBREG (v16f32 (VCVTUDQ2PSZrr (v16i32 (SUBREG_TO_REG (i32 0), VR128X:$src1, sub_xmm)))), sub_xmm)>; +def : Pat<(v4f64 (uint_to_fp (v4i32 VR128X:$src1))), + (EXTRACT_SUBREG (v8f64 (VCVTUDQ2PDZrr + (v8i32 (SUBREG_TO_REG (i32 0), VR128X:$src1, sub_xmm)))), sub_ymm)>; + def : Pat<(v16f32 (int_x86_avx512_mask_cvtdq2ps_512 (v16i32 VR512:$src), (bc_v16f32 (v16i32 immAllZerosV)), (i16 -1), imm:$rc)), (VCVTDQ2PSZrrb VR512:$src, imm:$rc)>; @@ -3452,7 +4237,6 @@ def : Pat <(v8f64 (int_x86_avx512_rcp28_pd (v8f64 VR512:$src), (VRCP28PDZrb VR512:$src)>; multiclass avx512_sqrt_packed opc, string OpcodeStr, SDNode OpNode, - Intrinsic V16F32Int, Intrinsic V8F64Int, OpndItins itins_s, OpndItins itins_d> { def PSZrr :AVX512PSI opc, string OpcodeStr, SDNode OpNode, (v8f64 (bitconvert (memopv16f32 addr:$src)))))], itins_d.rm>, EVEX, EVEX_V512, EVEX_CD8<64, CD8VF>; -let isCodeGenOnly = 1 in { - def PSZr_Int : AVX512PSI, - EVEX, EVEX_V512; - def PSZm_Int : AVX512PSI, EVEX, - EVEX_V512, EVEX_CD8<32, CD8VF>; - def PDZr_Int : AVX512PDI, - EVEX, EVEX_V512, VEX_W; - def PDZm_Int : AVX512PDI, - EVEX, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>; -} // isCodeGenOnly = 1 } multiclass avx512_sqrt_scalar opc, string OpcodeStr, @@ -3567,10 +4330,16 @@ defm VSQRT : avx512_sqrt_scalar<0x51, "sqrt", int_x86_avx512_sqrt_ss, int_x86_avx512_sqrt_sd, SSE_SQRTSS, SSE_SQRTSD>, avx512_sqrt_packed<0x51, "vsqrt", fsqrt, - int_x86_avx512_sqrt_ps_512, int_x86_avx512_sqrt_pd_512, SSE_SQRTPS, SSE_SQRTPD>; let Predicates = [HasAVX512] in { + def : Pat<(v16f32 (int_x86_avx512_sqrt_ps_512 (v16f32 VR512:$src1), + (bc_v16f32 (v16i32 immAllZerosV)), (i16 -1), FROUND_CURRENT)), + (VSQRTPSZrr VR512:$src1)>; + def : Pat<(v8f64 (int_x86_avx512_sqrt_pd_512 (v8f64 VR512:$src1), + (bc_v8f64 (v16i32 immAllZerosV)), (i8 -1), FROUND_CURRENT)), + (VSQRTPDZrr VR512:$src1)>; + def : Pat<(f32 (fsqrt FR32X:$src)), (VSQRTSSZr (f32 (IMPLICIT_DEF)), FR32X:$src)>; def : Pat<(f32 (fsqrt (load addr:$src))), @@ -3738,7 +4507,7 @@ defm VRNDSCALEPSZ : avx512_rndscale<0x08, "vrndscaleps", f512mem, VR512, EVEX_CD8<32, CD8VF>; def : Pat<(v16f32 (int_x86_avx512_mask_rndscale_ps_512 (v16f32 VR512:$src1), - imm:$src2, (bc_v16f32 (v16i32 immAllZerosV)), (i16 -1), + imm:$src2, (v16f32 VR512:$src1), (i16 -1), FROUND_CURRENT)), (VRNDSCALEPSZr VR512:$src1, imm:$src2)>; @@ -3748,7 +4517,7 @@ defm VRNDSCALEPDZ : avx512_rndscale<0x09, "vrndscalepd", f512mem, VR512, VEX_W, EVEX_CD8<64, CD8VF>; def : Pat<(v8f64 (int_x86_avx512_mask_rndscale_pd_512 (v8f64 VR512:$src1), - imm:$src2, (bc_v8f64 (v16i32 immAllZerosV)), (i8 -1), + imm:$src2, (v8f64 VR512:$src1), (i8 -1), FROUND_CURRENT)), (VRNDSCALEPDZr VR512:$src1, imm:$src2)>; @@ -3830,7 +4599,13 @@ multiclass avx512_trunc_sat opc, string OpcodeStr, !strconcat(OpcodeStr," \t{$src, $dst|$dst, $src}"), []>, EVEX; - def krr : AVX512XS8I, EVEX, EVEX_K; + + def rrkz : AVX512XS8I opc, string OpcodeStr, def mr : AVX512XS8I, EVEX; + + def mrk : AVX512XS8I, EVEX, EVEX_K; + } defm VPMOVQB : avx512_trunc_sat<0x32, "vpmovqb", VR128X, VR512, VK8WM, i128mem>, EVEX_V512, EVEX_CD8<8, CD8VO>; @@ -3878,60 +4659,86 @@ def : Pat<(v16i8 (X86vtrunc (v16i32 VR512:$src))), (VPMOVDBrr VR512:$src)>; def : Pat<(v8i32 (X86vtrunc (v8i64 VR512:$src))), (VPMOVQDrr VR512:$src)>; def : Pat<(v16i8 (X86vtruncm VK16WM:$mask, (v16i32 VR512:$src))), - (VPMOVDBkrr VK16WM:$mask, VR512:$src)>; + (VPMOVDBrrkz VK16WM:$mask, VR512:$src)>; def : Pat<(v16i16 (X86vtruncm VK16WM:$mask, (v16i32 VR512:$src))), - (VPMOVDWkrr VK16WM:$mask, VR512:$src)>; + (VPMOVDWrrkz VK16WM:$mask, VR512:$src)>; def : Pat<(v8i16 (X86vtruncm VK8WM:$mask, (v8i64 VR512:$src))), - (VPMOVQWkrr VK8WM:$mask, VR512:$src)>; + (VPMOVQWrrkz VK8WM:$mask, VR512:$src)>; def : Pat<(v8i32 (X86vtruncm VK8WM:$mask, (v8i64 VR512:$src))), - (VPMOVQDkrr VK8WM:$mask, VR512:$src)>; + (VPMOVQDrrkz VK8WM:$mask, VR512:$src)>; -multiclass avx512_extend opc, string OpcodeStr, RegisterClass DstRC, - RegisterClass SrcRC, SDNode OpNode, PatFrag mem_frag, - X86MemOperand x86memop, ValueType OpVT, ValueType InVT> { +multiclass avx512_extend opc, string OpcodeStr, RegisterClass KRC, + RegisterClass DstRC, RegisterClass SrcRC, SDNode OpNode, + PatFrag mem_frag, X86MemOperand x86memop, + ValueType OpVT, ValueType InVT> { def rr : AVX5128I, EVEX; - def rm : AVX5128I, EVEX, EVEX_K; + + def rrkz : AVX5128I, EVEX, EVEX_KZ; + + let mayLoad = 1 in { + def rm : AVX5128I, EVEX; + + def rmk : AVX5128I, + EVEX, EVEX_K; + + def rmkz : AVX5128I, + EVEX, EVEX_KZ; + } } -defm VPMOVZXBDZ: avx512_extend<0x31, "vpmovzxbd", VR512, VR128X, X86vzext, +defm VPMOVZXBDZ: avx512_extend<0x31, "vpmovzxbd", VK16WM, VR512, VR128X, X86vzext, memopv2i64, i128mem, v16i32, v16i8>, EVEX_V512, EVEX_CD8<8, CD8VQ>; -defm VPMOVZXBQZ: avx512_extend<0x32, "vpmovzxbq", VR512, VR128X, X86vzext, +defm VPMOVZXBQZ: avx512_extend<0x32, "vpmovzxbq", VK8WM, VR512, VR128X, X86vzext, memopv2i64, i128mem, v8i64, v16i8>, EVEX_V512, EVEX_CD8<8, CD8VO>; -defm VPMOVZXWDZ: avx512_extend<0x33, "vpmovzxwd", VR512, VR256X, X86vzext, +defm VPMOVZXWDZ: avx512_extend<0x33, "vpmovzxwd", VK16WM, VR512, VR256X, X86vzext, memopv4i64, i256mem, v16i32, v16i16>, EVEX_V512, EVEX_CD8<16, CD8VH>; -defm VPMOVZXWQZ: avx512_extend<0x34, "vpmovzxwq", VR512, VR128X, X86vzext, +defm VPMOVZXWQZ: avx512_extend<0x34, "vpmovzxwq", VK8WM, VR512, VR128X, X86vzext, memopv2i64, i128mem, v8i64, v8i16>, EVEX_V512, EVEX_CD8<16, CD8VQ>; -defm VPMOVZXDQZ: avx512_extend<0x35, "vpmovzxdq", VR512, VR256X, X86vzext, +defm VPMOVZXDQZ: avx512_extend<0x35, "vpmovzxdq", VK8WM, VR512, VR256X, X86vzext, memopv4i64, i256mem, v8i64, v8i32>, EVEX_V512, EVEX_CD8<32, CD8VH>; - -defm VPMOVSXBDZ: avx512_extend<0x21, "vpmovsxbd", VR512, VR128X, X86vsext, + +defm VPMOVSXBDZ: avx512_extend<0x21, "vpmovsxbd", VK16WM, VR512, VR128X, X86vsext, memopv2i64, i128mem, v16i32, v16i8>, EVEX_V512, EVEX_CD8<8, CD8VQ>; -defm VPMOVSXBQZ: avx512_extend<0x22, "vpmovsxbq", VR512, VR128X, X86vsext, +defm VPMOVSXBQZ: avx512_extend<0x22, "vpmovsxbq", VK8WM, VR512, VR128X, X86vsext, memopv2i64, i128mem, v8i64, v16i8>, EVEX_V512, EVEX_CD8<8, CD8VO>; -defm VPMOVSXWDZ: avx512_extend<0x23, "vpmovsxwd", VR512, VR256X, X86vsext, +defm VPMOVSXWDZ: avx512_extend<0x23, "vpmovsxwd", VK16WM, VR512, VR256X, X86vsext, memopv4i64, i256mem, v16i32, v16i16>, EVEX_V512, EVEX_CD8<16, CD8VH>; -defm VPMOVSXWQZ: avx512_extend<0x24, "vpmovsxwq", VR512, VR128X, X86vsext, +defm VPMOVSXWQZ: avx512_extend<0x24, "vpmovsxwq", VK8WM, VR512, VR128X, X86vsext, memopv2i64, i128mem, v8i64, v8i16>, EVEX_V512, EVEX_CD8<16, CD8VQ>; -defm VPMOVSXDQZ: avx512_extend<0x25, "vpmovsxdq", VR512, VR256X, X86vsext, +defm VPMOVSXDQZ: avx512_extend<0x25, "vpmovsxdq", VK8WM, VR512, VR256X, X86vsext, memopv4i64, i256mem, v8i64, v8i32>, EVEX_V512, EVEX_CD8<32, CD8VH>; @@ -4007,6 +4814,62 @@ defm VPSCATTERQQZ : avx512_scatter<0xA1, "vpscatterqq", VK8WM, VR512, vz64mem>, defm VPSCATTERQDZ : avx512_scatter<0xA1, "vpscatterqd", VK8WM, VR256X, vz64mem>, EVEX_V512, EVEX_CD8<32, CD8VT1>; +// prefetch +multiclass avx512_gather_scatter_prefetch opc, Format F, string OpcodeStr, + RegisterClass KRC, X86MemOperand memop> { + let Predicates = [HasPFI], hasSideEffects = 1 in + def m : AVX5128I, EVEX, EVEX_K; +} + +defm VGATHERPF0DPS: avx512_gather_scatter_prefetch<0xC6, MRM1m, "vgatherpf0dps", + VK16WM, vz32mem>, EVEX_V512, EVEX_CD8<32, CD8VT1>; + +defm VGATHERPF0QPS: avx512_gather_scatter_prefetch<0xC7, MRM1m, "vgatherpf0qps", + VK8WM, vz64mem>, EVEX_V512, EVEX_CD8<64, CD8VT1>; + +defm VGATHERPF0DPD: avx512_gather_scatter_prefetch<0xC6, MRM1m, "vgatherpf0dpd", + VK8WM, vy32mem>, EVEX_V512, VEX_W, EVEX_CD8<32, CD8VT1>; + +defm VGATHERPF0QPD: avx512_gather_scatter_prefetch<0xC7, MRM1m, "vgatherpf0qpd", + VK8WM, vz64mem>, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VT1>; + +defm VGATHERPF1DPS: avx512_gather_scatter_prefetch<0xC6, MRM2m, "vgatherpf1dps", + VK16WM, vz32mem>, EVEX_V512, EVEX_CD8<32, CD8VT1>; + +defm VGATHERPF1QPS: avx512_gather_scatter_prefetch<0xC7, MRM2m, "vgatherpf1qps", + VK8WM, vz64mem>, EVEX_V512, EVEX_CD8<64, CD8VT1>; + +defm VGATHERPF1DPD: avx512_gather_scatter_prefetch<0xC6, MRM2m, "vgatherpf1dpd", + VK8WM, vy32mem>, EVEX_V512, VEX_W, EVEX_CD8<32, CD8VT1>; + +defm VGATHERPF1QPD: avx512_gather_scatter_prefetch<0xC7, MRM2m, "vgatherpf1qpd", + VK8WM, vz64mem>, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VT1>; + +defm VSCATTERPF0DPS: avx512_gather_scatter_prefetch<0xC6, MRM5m, "vscatterpf0dps", + VK16WM, vz32mem>, EVEX_V512, EVEX_CD8<32, CD8VT1>; + +defm VSCATTERPF0QPS: avx512_gather_scatter_prefetch<0xC7, MRM5m, "vscatterpf0qps", + VK8WM, vz64mem>, EVEX_V512, EVEX_CD8<64, CD8VT1>; + +defm VSCATTERPF0DPD: avx512_gather_scatter_prefetch<0xC6, MRM5m, "vscatterpf0dpd", + VK8WM, vy32mem>, EVEX_V512, VEX_W, EVEX_CD8<32, CD8VT1>; + +defm VSCATTERPF0QPD: avx512_gather_scatter_prefetch<0xC7, MRM5m, "vscatterpf0qpd", + VK8WM, vz64mem>, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VT1>; + +defm VSCATTERPF1DPS: avx512_gather_scatter_prefetch<0xC6, MRM6m, "vscatterpf1dps", + VK16WM, vz32mem>, EVEX_V512, EVEX_CD8<32, CD8VT1>; + +defm VSCATTERPF1QPS: avx512_gather_scatter_prefetch<0xC7, MRM6m, "vscatterpf1qps", + VK8WM, vz64mem>, EVEX_V512, EVEX_CD8<64, CD8VT1>; + +defm VSCATTERPF1DPD: avx512_gather_scatter_prefetch<0xC6, MRM6m, "vscatterpf1dpd", + VK8WM, vy32mem>, EVEX_V512, VEX_W, EVEX_CD8<32, CD8VT1>; + +defm VSCATTERPF1QPD: avx512_gather_scatter_prefetch<0xC7, MRM6m, "vscatterpf1qpd", + VK8WM, vz64mem>, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VT1>; //===----------------------------------------------------------------------===// // VSHUFPS - VSHUFPD Operations @@ -4046,33 +4909,29 @@ def : Pat<(v8i64 (X86Shufp VR512:$src1, (memopv8i64 addr:$src2), (i8 imm:$imm))), (VSHUFPDZrmi VR512:$src1, addr:$src2, imm:$imm)>; -multiclass avx512_alignr { - def rri : AVX512AIi8<0x03, MRMSrcReg, (outs RC:$dst), - (ins RC:$src1, RC:$src2, i8imm:$src3), - !strconcat(OpcodeStr, - " \t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"), - []>, EVEX_4V; +multiclass avx512_valign { + defm rri : AVX512_maskable<0x03, MRMSrcReg, _, (outs _.RC:$dst), + (ins _.RC:$src1, _.RC:$src2, i8imm:$src3), + "valign"##_.Suffix, + "$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:$imm))), + (!cast(NAME##rri) _.RC:$src2, _.RC:$src1, imm:$imm)>; + let mayLoad = 1 in - def rmi : AVX512AIi8<0x03, MRMSrcMem, (outs RC:$dst), - (ins RC:$src1, x86memop:$src2, i8imm:$src3), - !strconcat(OpcodeStr, - " \t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"), + def rmi : AVX512AIi8<0x03, MRMSrcMem, (outs _.RC:$dst), + (ins _.RC:$src1, _.MemOp:$src2, i8imm:$src3), + !strconcat("valign"##_.Suffix, + " \t{$src3, $src2, $src1, $dst|" + "$dst, $src1, $src2, $src3}"), []>, EVEX_4V; } -defm VALIGND : avx512_alignr<"valignd", VR512, i512mem>, - EVEX_V512, EVEX_CD8<32, CD8VF>; -defm VALIGNQ : avx512_alignr<"valignq", VR512, i512mem>, - VEX_W, EVEX_V512, EVEX_CD8<64, CD8VF>; - -def : Pat<(v16f32 (X86PAlignr VR512:$src1, VR512:$src2, (i8 imm:$imm))), - (VALIGNDrri VR512:$src2, VR512:$src1, imm:$imm)>; -def : Pat<(v8f64 (X86PAlignr VR512:$src1, VR512:$src2, (i8 imm:$imm))), - (VALIGNQrri VR512:$src2, VR512:$src1, imm:$imm)>; -def : Pat<(v16i32 (X86PAlignr VR512:$src1, VR512:$src2, (i8 imm:$imm))), - (VALIGNDrri VR512:$src2, VR512:$src1, imm:$imm)>; -def : Pat<(v8i64 (X86PAlignr VR512:$src1, VR512:$src2, (i8 imm:$imm))), - (VALIGNQrri VR512:$src2, VR512:$src1, imm:$imm)>; +defm VALIGND : avx512_valign, EVEX_V512, EVEX_CD8<32, CD8VF>; +defm VALIGNQ : avx512_valign, VEX_W, EVEX_V512, EVEX_CD8<64, CD8VF>; // Helper fragments to match sext vXi1 to vXiY. def v16i1sextv16i32 : PatLeaf<(v16i32 (X86vsrai VR512:$src, (i8 31)))>; @@ -4224,6 +5083,78 @@ def : Pat<(int_x86_avx512_mask_conflict_q_512 VR512:$src2, VR512:$src1, (VPCONFLICTQrrk VR512:$src1, (v8i1 (COPY_TO_REGCLASS GR8:$mask, VK8WM)), VR512:$src2)>; +let Predicates = [HasCDI] in { +defm VPLZCNTD : avx512_conflict<0x44, "vplzcntd", VR512, VK16WM, + i512mem, i32mem, "{1to16}">, + EVEX_V512, EVEX_CD8<32, CD8VF>; + + +defm VPLZCNTQ : avx512_conflict<0x44, "vplzcntq", VR512, VK8WM, + i512mem, i64mem, "{1to8}">, + EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>; + +} + +def : Pat<(int_x86_avx512_mask_lzcnt_d_512 VR512:$src2, VR512:$src1, + GR16:$mask), + (VPLZCNTDrrk VR512:$src1, + (v16i1 (COPY_TO_REGCLASS GR16:$mask, VK16WM)), VR512:$src2)>; + +def : Pat<(int_x86_avx512_mask_lzcnt_q_512 VR512:$src2, VR512:$src1, + GR8:$mask), + (VPLZCNTQrrk VR512:$src1, + (v8i1 (COPY_TO_REGCLASS GR8:$mask, VK8WM)), VR512:$src2)>; + +def : Pat<(v16i32 (ctlz (memopv16i32 addr:$src))), + (VPLZCNTDrm addr:$src)>; +def : Pat<(v16i32 (ctlz (v16i32 VR512:$src))), + (VPLZCNTDrr VR512:$src)>; +def : Pat<(v8i64 (ctlz (memopv8i64 addr:$src))), + (VPLZCNTQrm addr:$src)>; +def : Pat<(v8i64 (ctlz (v8i64 VR512:$src))), + (VPLZCNTQrr VR512:$src)>; + def : Pat<(store (i1 -1), addr:$dst), (MOV8mi addr:$dst, (i8 1))>; def : Pat<(store (i1 1), addr:$dst), (MOV8mi addr:$dst, (i8 1))>; def : Pat<(store (i1 0), addr:$dst), (MOV8mi addr:$dst, (i8 0))>; + +def : Pat<(store VK1:$src, addr:$dst), + (KMOVWmk addr:$dst, (COPY_TO_REGCLASS VK1:$src, VK16))>; + +def truncstorei1 : PatFrag<(ops node:$val, node:$ptr), + (truncstore node:$val, node:$ptr), [{ + return cast(N)->getMemoryVT() == MVT::i1; +}]>; + +def : Pat<(truncstorei1 GR8:$src, addr:$dst), + (MOV8mr addr:$dst, GR8:$src)>; + +multiclass cvt_by_vec_width opc, X86VectorVTInfo Vec, string OpcodeStr > { +def rr : AVX512XS8I, EVEX; +} + +multiclass cvt_mask_by_elt_width opc, AVX512VLVectorVTInfo VTInfo, + string OpcodeStr, Predicate prd> { +let Predicates = [prd] in + defm Z : cvt_by_vec_width, EVEX_V512; + + let Predicates = [prd, HasVLX] in { + defm Z256 : cvt_by_vec_width, EVEX_V256; + defm Z128 : cvt_by_vec_width, EVEX_V128; + } +} + +multiclass avx512_convert_mask_to_vector { + defm NAME##B : cvt_mask_by_elt_width<0x28, avx512vl_i8_info, OpcodeStr, + HasBWI>; + defm NAME##W : cvt_mask_by_elt_width<0x28, avx512vl_i16_info, OpcodeStr, + HasBWI>, VEX_W; + defm NAME##D : cvt_mask_by_elt_width<0x38, avx512vl_i32_info, OpcodeStr, + HasDQI>; + defm NAME##Q : cvt_mask_by_elt_width<0x38, avx512vl_i64_info, OpcodeStr, + HasDQI>, VEX_W; +} + +defm VPMOVM2 : avx512_convert_mask_to_vector<"vpmovm2">;