================ @@ -0,0 +1,588 @@ +//===-- X86InstrAVX10_V2_AUX.td - AVX10 V2 AUX Instructions --*- tablegen -*-===// +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// +//===----------------------------------------------------------------------===// +// +// This file describes the X86 AVX10 V2 AUX instruction set, defining the +// instructions and their encoding. +// +//===----------------------------------------------------------------------===// + +//===----------------------------------------------------------------------===// +// AVX10 V2 AUX Multiclass Definitions +//===----------------------------------------------------------------------===// + +// Group A multiclass: PS(f32) -> i8 truncating conversion (quarter-size output) +// Output is always xmm for all VL variants. +// Adapts avx512_vcvt_fp for each VL level with explicit dest/src type mappings. +multiclass avx10_v2aux_cvt_trunc_ps2i8<bits<8> opc, string OpcodeStr, + SDPatternOperator OpNode, + SDPatternOperator MaskOpNode> { + let Predicates = [HasAVX10_V2_AUX] in { + let ExeDomain = SSEPackedSingle in { + let Uses = []<Register>, mayRaiseFPException = 0 in { + defm Z : avx512_vcvt_fp<opc, OpcodeStr, v16i8x_info, v16f32_info, + OpNode, OpNode, WriteCvtPH2PSZ>, EVEX_V512; + // Z256/Z128: use null_frag because element count mismatch between + // dest (v16i8) and source (v8f32/v4f32) prevents avx512_vcvt_fp from + // generating correct masked patterns. Explicit Pat patterns below. + defm Z256 : avx512_vcvt_fp<opc, OpcodeStr, v16i8x_info, v8f32x_info, + null_frag, null_frag, + WriteCvtPH2PSZ, v8f32x_info.BroadcastStr, + "{y}", v8f32x_info.MemOp, + v8f32x_info.KRCWM>, EVEX_V256; + defm Z128 : avx512_vcvt_fp<opc, OpcodeStr, v16i8x_info, v4f32x_info, + null_frag, null_frag, + WriteCvtPH2PSZ, v4f32x_info.BroadcastStr, + "{x}", f128mem, + v4f32x_info.KRCWM>, EVEX_V128; + } + } + + // InstAliases for x/y suffixes + def : InstAlias<OpcodeStr#"x\t{$src, $dst|$dst, $src}", + (!cast<Instruction>(NAME # "Z128rr") VR128X:$dst, + VR128X:$src), 0>; + def : InstAlias<OpcodeStr#"x\t{$src, $dst|$dst, $src}", + (!cast<Instruction>(NAME # "Z128rm") VR128X:$dst, + f128mem:$src), 0, "intel">; + def : InstAlias<OpcodeStr#"y\t{$src, $dst|$dst, $src}", + (!cast<Instruction>(NAME # "Z256rr") VR128X:$dst, + VR256X:$src), 0>; + def : InstAlias<OpcodeStr#"y\t{$src, $dst|$dst, $src}", + (!cast<Instruction>(NAME # "Z256rm") VR128X:$dst, + f256mem:$src), 0, "intel">; + + // Explicit patterns for Z256 (8 source elements, VK8WM mask) + // Unmasked + def : Pat<(v16i8 (OpNode (v8f32 VR256X:$src))), + (!cast<Instruction>(NAME # "Z256rr") VR256X:$src)>; + // Masked (merge) + def : Pat<(MaskOpNode (v8f32 VR256X:$src), (v16i8 VR128X:$src0), + VK8WM:$mask), + (!cast<Instruction>(NAME # "Z256rrk") VR128X:$src0, VK8WM:$mask, + VR256X:$src)>; + // Masked (zero) + def : Pat<(MaskOpNode (v8f32 VR256X:$src), v16i8x_info.ImmAllZerosV, + VK8WM:$mask), + (!cast<Instruction>(NAME # "Z256rrkz") VK8WM:$mask, + VR256X:$src)>; + // Memory + def : Pat<(v16i8 (OpNode (loadv8f32 addr:$src))), + (!cast<Instruction>(NAME # "Z256rm") addr:$src)>; + def : Pat<(MaskOpNode (loadv8f32 addr:$src), (v16i8 VR128X:$src0), + VK8WM:$mask), + (!cast<Instruction>(NAME # "Z256rmk") VR128X:$src0, VK8WM:$mask, + addr:$src)>; + def : Pat<(MaskOpNode (loadv8f32 addr:$src), v16i8x_info.ImmAllZerosV, + VK8WM:$mask), + (!cast<Instruction>(NAME # "Z256rmkz") VK8WM:$mask, addr:$src)>; + // Broadcast + def : Pat<(v16i8 (OpNode (v8f32 (X86VBroadcastld32 addr:$src)))), + (!cast<Instruction>(NAME # "Z256rmb") addr:$src)>; + def : Pat<(MaskOpNode (v8f32 (X86VBroadcastld32 addr:$src)), + (v16i8 VR128X:$src0), VK8WM:$mask), + (!cast<Instruction>(NAME # "Z256rmbk") VR128X:$src0, VK8WM:$mask, + addr:$src)>; + def : Pat<(MaskOpNode (v8f32 (X86VBroadcastld32 addr:$src)), + v16i8x_info.ImmAllZerosV, VK8WM:$mask), + (!cast<Instruction>(NAME # "Z256rmbkz") VK8WM:$mask, addr:$src)>; + + // Explicit patterns for Z128 (4 source elements, VK4WM mask) + // Unmasked + def : Pat<(v16i8 (OpNode (v4f32 VR128X:$src))), + (!cast<Instruction>(NAME # "Z128rr") VR128X:$src)>; + // Masked (merge) + def : Pat<(MaskOpNode (v4f32 VR128X:$src), (v16i8 VR128X:$src0), + VK4WM:$mask), + (!cast<Instruction>(NAME # "Z128rrk") VR128X:$src0, VK4WM:$mask, + VR128X:$src)>; + // Masked (zero) + def : Pat<(MaskOpNode (v4f32 VR128X:$src), v16i8x_info.ImmAllZerosV, + VK4WM:$mask), + (!cast<Instruction>(NAME # "Z128rrkz") VK4WM:$mask, + VR128X:$src)>; + // Memory + def : Pat<(v16i8 (OpNode (loadv4f32 addr:$src))), + (!cast<Instruction>(NAME # "Z128rm") addr:$src)>; + def : Pat<(MaskOpNode (loadv4f32 addr:$src), (v16i8 VR128X:$src0), + VK4WM:$mask), + (!cast<Instruction>(NAME # "Z128rmk") VR128X:$src0, VK4WM:$mask, + addr:$src)>; + def : Pat<(MaskOpNode (loadv4f32 addr:$src), v16i8x_info.ImmAllZerosV, + VK4WM:$mask), + (!cast<Instruction>(NAME # "Z128rmkz") VK4WM:$mask, addr:$src)>; + // Broadcast + def : Pat<(v16i8 (OpNode (v4f32 (X86VBroadcastld32 addr:$src)))), + (!cast<Instruction>(NAME # "Z128rmb") addr:$src)>; + def : Pat<(MaskOpNode (v4f32 (X86VBroadcastld32 addr:$src)), + (v16i8 VR128X:$src0), VK4WM:$mask), + (!cast<Instruction>(NAME # "Z128rmbk") VR128X:$src0, VK4WM:$mask, + addr:$src)>; + def : Pat<(MaskOpNode (v4f32 (X86VBroadcastld32 addr:$src)), + v16i8x_info.ImmAllZerosV, VK4WM:$mask), + (!cast<Instruction>(NAME # "Z128rmbkz") VK4WM:$mask, addr:$src)>; + } +} + +// Group B multiclass: 3-operand bias PS(f32) -> i8 conversion (quarter-size output) +// bias + f32 source -> i8 dest +// Reuses avx10_convert_3op_packed from X86InstrAVX10.td for each VL variant. +multiclass avx10_v2aux_convert_3op_ps<bits<8> OpCode, string OpcodeStr, + SDPatternOperator OpNode, + SDPatternOperator MaskOpNode> { + let Predicates = [HasAVX10_V2_AUX] in { + // Z (512-bit): bias=v64i8(zmm), src=v16f32(zmm), dst=v16i8(xmm) + // Element counts match (16), so vselect_mask works directly. + defm Z : avx10_convert_3op_packed<OpCode, OpcodeStr, v16i8x_info, + v64i8_info, v16f32_info, OpNode, OpNode, WriteCvtPH2PSZ>, + EVEX_V512, EVEX_CD8<32, CD8VF>; + // Z256/Z128: use null_frag because element count mismatch between + // dest (v16i8) and source (v8f32/v4f32) prevents vselect_mask from + // generating correct masked patterns. Explicit Pat patterns below. + defm Z256 : avx10_convert_3op_packed<OpCode, OpcodeStr, v16i8x_info, + v32i8x_info, v8f32x_info, + null_frag, null_frag, WriteCvtPH2PSZ>, + EVEX_V256, EVEX_CD8<32, CD8VF>; + defm Z128 : avx10_convert_3op_packed<OpCode, OpcodeStr, v16i8x_info, + v16i8x_info, v4f32x_info, + null_frag, null_frag, WriteCvtPH2PSZ>, + EVEX_V128, EVEX_CD8<32, CD8VF>; + + // Explicit patterns for Z256 (8 source elements, VK8WM mask) + def : Pat<(v16i8 (OpNode (v32i8 VR256X:$src1), (v8f32 VR256X:$src2))), + (!cast<Instruction>(NAME # "Z256rr") VR256X:$src1, VR256X:$src2)>; + def : Pat<(MaskOpNode (v32i8 VR256X:$src1), (v8f32 VR256X:$src2), + (v16i8 VR128X:$src0), VK8WM:$mask), + (!cast<Instruction>(NAME # "Z256rrk") VR128X:$src0, VK8WM:$mask, + VR256X:$src1, VR256X:$src2)>; + def : Pat<(MaskOpNode (v32i8 VR256X:$src1), (v8f32 VR256X:$src2), + v16i8x_info.ImmAllZerosV, VK8WM:$mask), + (!cast<Instruction>(NAME # "Z256rrkz") VK8WM:$mask, + VR256X:$src1, VR256X:$src2)>; + // Memory + def : Pat<(v16i8 (OpNode (v32i8 VR256X:$src1), (loadv8f32 addr:$src2))), + (!cast<Instruction>(NAME # "Z256rm") VR256X:$src1, addr:$src2)>; + def : Pat<(MaskOpNode (v32i8 VR256X:$src1), (loadv8f32 addr:$src2), + (v16i8 VR128X:$src0), VK8WM:$mask), + (!cast<Instruction>(NAME # "Z256rmk") VR128X:$src0, VK8WM:$mask, + VR256X:$src1, addr:$src2)>; + def : Pat<(MaskOpNode (v32i8 VR256X:$src1), (loadv8f32 addr:$src2), + v16i8x_info.ImmAllZerosV, VK8WM:$mask), + (!cast<Instruction>(NAME # "Z256rmkz") VK8WM:$mask, + VR256X:$src1, addr:$src2)>; + // Broadcast + def : Pat<(v16i8 (OpNode (v32i8 VR256X:$src1), + (v8f32 (X86VBroadcastld32 addr:$src2)))), + (!cast<Instruction>(NAME # "Z256rmb") VR256X:$src1, addr:$src2)>; + def : Pat<(MaskOpNode (v32i8 VR256X:$src1), + (v8f32 (X86VBroadcastld32 addr:$src2)), + (v16i8 VR128X:$src0), VK8WM:$mask), + (!cast<Instruction>(NAME # "Z256rmbk") VR128X:$src0, VK8WM:$mask, + VR256X:$src1, addr:$src2)>; + def : Pat<(MaskOpNode (v32i8 VR256X:$src1), + (v8f32 (X86VBroadcastld32 addr:$src2)), + v16i8x_info.ImmAllZerosV, VK8WM:$mask), + (!cast<Instruction>(NAME # "Z256rmbkz") VK8WM:$mask, + VR256X:$src1, addr:$src2)>; + + // Explicit patterns for Z128 (4 source elements, VK4WM mask) + def : Pat<(v16i8 (OpNode (v16i8 VR128X:$src1), (v4f32 VR128X:$src2))), + (!cast<Instruction>(NAME # "Z128rr") VR128X:$src1, VR128X:$src2)>; + def : Pat<(MaskOpNode (v16i8 VR128X:$src1), (v4f32 VR128X:$src2), + (v16i8 VR128X:$src0), VK4WM:$mask), + (!cast<Instruction>(NAME # "Z128rrk") VR128X:$src0, VK4WM:$mask, + VR128X:$src1, VR128X:$src2)>; + def : Pat<(MaskOpNode (v16i8 VR128X:$src1), (v4f32 VR128X:$src2), + v16i8x_info.ImmAllZerosV, VK4WM:$mask), + (!cast<Instruction>(NAME # "Z128rrkz") VK4WM:$mask, + VR128X:$src1, VR128X:$src2)>; + // Memory + def : Pat<(v16i8 (OpNode (v16i8 VR128X:$src1), (loadv4f32 addr:$src2))), + (!cast<Instruction>(NAME # "Z128rm") VR128X:$src1, addr:$src2)>; + def : Pat<(MaskOpNode (v16i8 VR128X:$src1), (loadv4f32 addr:$src2), + (v16i8 VR128X:$src0), VK4WM:$mask), + (!cast<Instruction>(NAME # "Z128rmk") VR128X:$src0, VK4WM:$mask, + VR128X:$src1, addr:$src2)>; + def : Pat<(MaskOpNode (v16i8 VR128X:$src1), (loadv4f32 addr:$src2), + v16i8x_info.ImmAllZerosV, VK4WM:$mask), + (!cast<Instruction>(NAME # "Z128rmkz") VK4WM:$mask, + VR128X:$src1, addr:$src2)>; + // Broadcast + def : Pat<(v16i8 (OpNode (v16i8 VR128X:$src1), + (v4f32 (X86VBroadcastld32 addr:$src2)))), + (!cast<Instruction>(NAME # "Z128rmb") VR128X:$src1, addr:$src2)>; + def : Pat<(MaskOpNode (v16i8 VR128X:$src1), + (v4f32 (X86VBroadcastld32 addr:$src2)), + (v16i8 VR128X:$src0), VK4WM:$mask), + (!cast<Instruction>(NAME # "Z128rmbk") VR128X:$src0, VK4WM:$mask, + VR128X:$src1, addr:$src2)>; + def : Pat<(MaskOpNode (v16i8 VR128X:$src1), + (v4f32 (X86VBroadcastld32 addr:$src2)), + v16i8x_info.ImmAllZerosV, VK4WM:$mask), + (!cast<Instruction>(NAME # "Z128rmbkz") VK4WM:$mask, + VR128X:$src1, addr:$src2)>; + } +} + +// Group C multiclass: i8 -> f32 expanding conversion (4x expansion, no broadcast) +multiclass avx10_v2aux_convert_2op_i8_to_f32<string OpcodeStr, bits<8> opc, + SDNode OpNode> { + let Predicates = [HasAVX10_V2_AUX] in { + defm Z : avx10_convert_2op_nomb_packed<opc, OpcodeStr, v16f32_info, + v16i8x_info, OpNode, f128mem, + WriteCvtPH2PSZ>, EVEX_V512; + defm Z128 : avx10_convert_2op_nomb_packed<opc, OpcodeStr, v4f32x_info, + v16i8x_info, OpNode, f32mem, + WriteCvtPH2PSZ>, EVEX_V128; + defm Z256 : avx10_convert_2op_nomb_packed<opc, OpcodeStr, v8f32x_info, + v16i8x_info, OpNode, f64mem, + WriteCvtPH2PSZ>, EVEX_V256; + } +} + +// Group D multiclass: Store-like truncation (reg/mem dest, no masking) +// Uses MRMDestReg/MRMDestMem since destination can be memory operand. +// Source is in reg field, destination is in r/m field. +multiclass avx10_v2aux_trunc_store<bits<8> opc, string OpcodeStr, + X86VectorVTInfo SrcInfo, + X86VectorVTInfo DestInfo, + X86MemOperand x86memop> { + let hasSideEffects = 0 in { + def rr : I<opc, MRMDestReg, (outs DestInfo.RC:$dst), + (ins SrcInfo.RC:$src), + OpcodeStr # "\t{$src, $dst|$dst, $src}", []>, + EVEX, Sched<[WriteCvtPH2PSZ]>; + let mayStore = 1 in + def mr : I<opc, MRMDestMem, (outs), + (ins x86memop:$dst, SrcInfo.RC:$src), + OpcodeStr # "\t{$src, $dst|$dst, $src}", []>, + EVEX, Sched<[WriteCvtPH2PSZ.Folded]>; + } +} + +// Group F helper: expanding conversion with masking, no broadcast (reg+mem) +multiclass avx10_v2aux_convert_expand_masked<bits<8> opc, string OpcodeStr, + X86VectorVTInfo _dest, + X86VectorVTInfo _src, + X86MemOperand x86memop> { + let ExeDomain = _dest.ExeDomain in { + defm rr : AVX512_maskable_custom<opc, MRMSrcReg, + (outs _dest.RC:$dst), + (ins _src.RC:$src), + (ins _dest.RC:$src0, _dest.KRCWM:$mask, _src.RC:$src), + (ins _dest.KRCWM:$mask, _src.RC:$src), + OpcodeStr, "$src", "$src", [], [], [], + "$src0 = $dst">, + EVEX, Sched<[WriteCvtPH2PSZ]>; + let mayLoad = 1 in + defm rm : AVX512_maskable_custom<opc, MRMSrcMem, + (outs _dest.RC:$dst), + (ins x86memop:$src), + (ins _dest.RC:$src0, _dest.KRCWM:$mask, x86memop:$src), + (ins _dest.KRCWM:$mask, x86memop:$src), + OpcodeStr, "$src", "$src", [], [], [], + "$src0 = $dst">, + EVEX, Sched<[WriteCvtPH2PSZ.Folded]>; + } +} + +// Group F helper: same-size conversion with masking (reg-only) +multiclass avx10_v2aux_convert_samesize_masked<bits<8> opc, string OpcodeStr, ---------------- mahesh-attarde wrote:
using "samesize" raises ambiguity. consider BF6 -> HF8 is same size but BF4 ->HF8 or BF8->BF6 is not ? In general, we will need to extract more intuitive names for this hierarchy as we make progress in this spec. https://github.com/llvm/llvm-project/pull/206888 _______________________________________________ cfe-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
