diff options
Diffstat (limited to 'llvm/lib/Target/X86/X86InstrAVX512.td')
| -rw-r--r-- | llvm/lib/Target/X86/X86InstrAVX512.td | 308 |
1 files changed, 182 insertions, 126 deletions
diff --git a/llvm/lib/Target/X86/X86InstrAVX512.td b/llvm/lib/Target/X86/X86InstrAVX512.td index d0985f112b8..ccbe1f41912 100644 --- a/llvm/lib/Target/X86/X86InstrAVX512.td +++ b/llvm/lib/Target/X86/X86InstrAVX512.td @@ -145,6 +145,8 @@ def v2f64x_info : X86VectorVTInfo<2, f64, VR128X, "pd">; // We map scalar types to the smallest (128-bit) vector type // with the appropriate element type. This allows to use the same masking logic. +def i32x_info : X86VectorVTInfo<1, i32, GR32, "si">; +def i64x_info : X86VectorVTInfo<1, i64, GR64, "sq">; def f32x_info : X86VectorVTInfo<1, f32, VR128X, "ss">; def f64x_info : X86VectorVTInfo<1, f64, VR128X, "sd">; @@ -4598,50 +4600,55 @@ def : Pat<(f64 (uint_to_fp GR64:$src)), //===----------------------------------------------------------------------===// // AVX-512 Scalar convert from float/double to integer //===----------------------------------------------------------------------===// -multiclass avx512_cvt_s_int<bits<8> opc, RegisterClass SrcRC, RegisterClass DstRC, - Intrinsic Int, Operand memop, ComplexPattern mem_cpat, - string asm> { -let hasSideEffects = 0 in { - def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src), - !strconcat(asm,"\t{$src, $dst|$dst, $src}"), - [(set DstRC:$dst, (Int SrcRC:$src))]>, EVEX, VEX_LIG, - Requires<[HasAVX512]>; - let mayLoad = 1 in - def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst), (ins memop:$src), - !strconcat(asm,"\t{$src, $dst|$dst, $src}"), []>, EVEX, VEX_LIG, - Requires<[HasAVX512]>; -} // hasSideEffects = 0 +multiclass avx512_cvt_s_int_round<bits<8> opc, RegisterClass SrcRC, + RegisterClass DstRC, Intrinsic Int, + Operand memop, ComplexPattern mem_cpat, string asm> { + let hasSideEffects = 0, Predicates = [HasAVX512] in { + def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src), + !strconcat(asm,"\t{$src, $dst|$dst, $src}"), + [(set DstRC:$dst, (Int SrcRC:$src))]>, EVEX, VEX_LIG; + def rb : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src, AVX512RC:$rc), + !strconcat(asm,"\t{$rc, $src, $dst|$dst, $src, $rc}"), []>, + EVEX, VEX_LIG, EVEX_B, EVEX_RC; + let mayLoad = 1 in + def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst), (ins memop:$src), + !strconcat(asm,"\t{$src, $dst|$dst, $src}"), []>, EVEX, VEX_LIG; + } // hasSideEffects = 0, Predicates = [HasAVX512] } -let Predicates = [HasAVX512] in { + // Convert float/double to signed/unsigned int 32/64 -defm VCVTSS2SIZ: avx512_cvt_s_int<0x2D, VR128X, GR32, int_x86_sse_cvtss2si, +defm VCVTSS2SIZ: avx512_cvt_s_int_round<0x2D, VR128X, GR32, int_x86_sse_cvtss2si, ssmem, sse_load_f32, "cvtss2si">, XS, EVEX_CD8<32, CD8VT1>; -defm VCVTSS2SI64Z: avx512_cvt_s_int<0x2D, VR128X, GR64, int_x86_sse_cvtss2si64, +defm VCVTSS2SI64Z: avx512_cvt_s_int_round<0x2D, VR128X, GR64, + int_x86_sse_cvtss2si64, ssmem, sse_load_f32, "cvtss2si">, XS, VEX_W, EVEX_CD8<32, CD8VT1>; -defm VCVTSS2USIZ: avx512_cvt_s_int<0x79, VR128X, GR32, int_x86_avx512_cvtss2usi, +defm VCVTSS2USIZ: avx512_cvt_s_int_round<0x79, VR128X, GR32, + int_x86_avx512_cvtss2usi, ssmem, sse_load_f32, "cvtss2usi">, XS, EVEX_CD8<32, CD8VT1>; -defm VCVTSS2USI64Z: avx512_cvt_s_int<0x79, VR128X, GR64, +defm VCVTSS2USI64Z: avx512_cvt_s_int_round<0x79, VR128X, GR64, int_x86_avx512_cvtss2usi64, ssmem, sse_load_f32, "cvtss2usi">, XS, VEX_W, EVEX_CD8<32, CD8VT1>; -defm VCVTSD2SIZ: avx512_cvt_s_int<0x2D, VR128X, GR32, int_x86_sse2_cvtsd2si, +defm VCVTSD2SIZ: avx512_cvt_s_int_round<0x2D, VR128X, GR32, int_x86_sse2_cvtsd2si, sdmem, sse_load_f64, "cvtsd2si">, XD, EVEX_CD8<64, CD8VT1>; -defm VCVTSD2SI64Z: avx512_cvt_s_int<0x2D, VR128X, GR64, int_x86_sse2_cvtsd2si64, +defm VCVTSD2SI64Z: avx512_cvt_s_int_round<0x2D, VR128X, GR64, + int_x86_sse2_cvtsd2si64, sdmem, sse_load_f64, "cvtsd2si">, XD, VEX_W, EVEX_CD8<64, CD8VT1>; -defm VCVTSD2USIZ: avx512_cvt_s_int<0x79, VR128X, GR32, int_x86_avx512_cvtsd2usi, +defm VCVTSD2USIZ: avx512_cvt_s_int_round<0x79, VR128X, GR32, + int_x86_avx512_cvtsd2usi, sdmem, sse_load_f64, "cvtsd2usi">, XD, EVEX_CD8<64, CD8VT1>; -defm VCVTSD2USI64Z: avx512_cvt_s_int<0x79, VR128X, GR64, +defm VCVTSD2USI64Z: avx512_cvt_s_int_round<0x79, VR128X, GR64, int_x86_avx512_cvtsd2usi64, sdmem, sse_load_f64, "cvtsd2usi">, XD, VEX_W, EVEX_CD8<64, CD8VT1>; -let isCodeGenOnly = 1 in { +let isCodeGenOnly = 1 , Predicates = [HasAVX512] in { defm Int_VCVTSI2SSZ : sse12_cvt_sint_3addr<0x2A, GR32, VR128X, int_x86_sse_cvtsi2ss, i32mem, loadi32, "cvtsi2ss{l}", SSE_CVT_Scalar, 0>, XS, EVEX_4V; @@ -4658,121 +4665,170 @@ let isCodeGenOnly = 1 in { defm Int_VCVTUSI2SDZ : sse12_cvt_sint_3addr<0x2A, GR32, VR128X, int_x86_avx512_cvtusi2sd, i32mem, loadi32, "cvtusi2sd{l}", SSE_CVT_Scalar, 0>, XD, EVEX_4V; -} // isCodeGenOnly = 1 +} // isCodeGenOnly = 1, Predicates = [HasAVX512] // Convert float/double to signed/unsigned int 32/64 with truncation -let isCodeGenOnly = 1 in { - defm Int_VCVTTSS2SIZ : avx512_cvt_s_int<0x2C, VR128X, GR32, int_x86_sse_cvttss2si, - ssmem, sse_load_f32, "cvttss2si">, - XS, EVEX_CD8<32, CD8VT1>; - defm Int_VCVTTSS2SI64Z : avx512_cvt_s_int<0x2C, VR128X, GR64, - int_x86_sse_cvttss2si64, ssmem, sse_load_f32, - "cvttss2si">, XS, VEX_W, - EVEX_CD8<32, CD8VT1>; - defm Int_VCVTTSD2SIZ : avx512_cvt_s_int<0x2C, VR128X, GR32, int_x86_sse2_cvttsd2si, - sdmem, sse_load_f64, "cvttsd2si">, XD, - EVEX_CD8<64, CD8VT1>; - defm Int_VCVTTSD2SI64Z : avx512_cvt_s_int<0x2C, VR128X, GR64, - int_x86_sse2_cvttsd2si64, sdmem, sse_load_f64, - "cvttsd2si">, XD, VEX_W, - EVEX_CD8<64, CD8VT1>; - defm Int_VCVTTSS2USIZ : avx512_cvt_s_int<0x78, VR128X, GR32, - int_x86_avx512_cvttss2usi, ssmem, sse_load_f32, - "cvttss2usi">, XS, EVEX_CD8<32, CD8VT1>; - defm Int_VCVTTSS2USI64Z : avx512_cvt_s_int<0x78, VR128X, GR64, - int_x86_avx512_cvttss2usi64, ssmem, - sse_load_f32, "cvttss2usi">, XS, VEX_W, - EVEX_CD8<32, CD8VT1>; - defm Int_VCVTTSD2USIZ : avx512_cvt_s_int<0x78, VR128X, GR32, - int_x86_avx512_cvttsd2usi, - sdmem, sse_load_f64, "cvttsd2usi">, XD, - EVEX_CD8<64, CD8VT1>; - defm Int_VCVTTSD2USI64Z : avx512_cvt_s_int<0x78, VR128X, GR64, - int_x86_avx512_cvttsd2usi64, sdmem, - sse_load_f64, "cvttsd2usi">, XD, VEX_W, - EVEX_CD8<64, CD8VT1>; -} // isCodeGenOnly = 1 - -multiclass avx512_cvt_s<bits<8> opc, RegisterClass SrcRC, RegisterClass DstRC, - SDNode OpNode, X86MemOperand x86memop, PatFrag ld_frag, - string asm> { - def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src), +multiclass avx512_cvt_s_all<bits<8> opc, string asm, X86VectorVTInfo _SrcRC, + X86VectorVTInfo _DstRC, SDNode OpNode, + SDNode OpNodeRnd>{ +let Predicates = [HasAVX512] in { + def rr : SI<opc, MRMSrcReg, (outs _DstRC.RC:$dst), (ins _SrcRC.FRC:$src), !strconcat(asm,"\t{$src, $dst|$dst, $src}"), - [(set DstRC:$dst, (OpNode SrcRC:$src))]>, EVEX; - def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst), (ins x86memop:$src), + [(set _DstRC.RC:$dst, (OpNode _SrcRC.FRC:$src))]>, EVEX; + def rb : SI<opc, MRMSrcReg, (outs _DstRC.RC:$dst), (ins _SrcRC.FRC:$src), + !strconcat(asm,"\t{{sae}, $src, $dst|$dst, $src, {sae}}"), + []>, EVEX, EVEX_B; + def rm : SI<opc, MRMSrcMem, (outs _DstRC.RC:$dst), (ins _SrcRC.MemOp:$src), !strconcat(asm,"\t{$src, $dst|$dst, $src}"), - [(set DstRC:$dst, (OpNode (ld_frag addr:$src)))]>, EVEX; -} - -defm VCVTTSS2SIZ : avx512_cvt_s<0x2C, FR32X, GR32, fp_to_sint, f32mem, - loadf32, "cvttss2si">, XS, - EVEX_CD8<32, CD8VT1>; -defm VCVTTSS2USIZ : avx512_cvt_s<0x78, FR32X, GR32, fp_to_uint, f32mem, - loadf32, "cvttss2usi">, XS, - EVEX_CD8<32, CD8VT1>; -defm VCVTTSS2SI64Z : avx512_cvt_s<0x2C, FR32X, GR64, fp_to_sint, f32mem, - loadf32, "cvttss2si">, XS, VEX_W, - EVEX_CD8<32, CD8VT1>; -defm VCVTTSS2USI64Z : avx512_cvt_s<0x78, FR32X, GR64, fp_to_uint, f32mem, - loadf32, "cvttss2usi">, XS, VEX_W, - EVEX_CD8<32, CD8VT1>; -defm VCVTTSD2SIZ : avx512_cvt_s<0x2C, FR64X, GR32, fp_to_sint, f64mem, - loadf64, "cvttsd2si">, XD, - EVEX_CD8<64, CD8VT1>; -defm VCVTTSD2USIZ : avx512_cvt_s<0x78, FR64X, GR32, fp_to_uint, f64mem, - loadf64, "cvttsd2usi">, XD, - EVEX_CD8<64, CD8VT1>; -defm VCVTTSD2SI64Z : avx512_cvt_s<0x2C, FR64X, GR64, fp_to_sint, f64mem, - loadf64, "cvttsd2si">, XD, VEX_W, - EVEX_CD8<64, CD8VT1>; -defm VCVTTSD2USI64Z : avx512_cvt_s<0x78, FR64X, GR64, fp_to_uint, f64mem, - loadf64, "cvttsd2usi">, XD, VEX_W, - EVEX_CD8<64, CD8VT1>; + [(set _DstRC.RC:$dst, (OpNode (_SrcRC.ScalarLdFrag addr:$src)))]>, + EVEX; + + let isCodeGenOnly = 1,hasSideEffects = 0 in { + def rr_Int : SI<opc, MRMSrcReg, (outs _DstRC.RC:$dst), (ins _SrcRC.RC:$src), + !strconcat(asm,"\t{$src, $dst|$dst, $src}"), + [(set _DstRC.RC:$dst, (OpNodeRnd _SrcRC.RC:$src, + (i32 FROUND_CURRENT)))]>, EVEX, VEX_LIG; + def rb_Int : SI<opc, MRMSrcReg, (outs _DstRC.RC:$dst), (ins _SrcRC.RC:$src), + !strconcat(asm,"\t{{sae}, $src, $dst|$dst, $src, {sae}}"), + [(set _DstRC.RC:$dst, (OpNodeRnd _SrcRC.RC:$src, + (i32 FROUND_NO_EXC)))]>, + EVEX,VEX_LIG , EVEX_B; + let mayLoad = 1 in + def rm_Int : SI<opc, MRMSrcMem, (outs _DstRC.RC:$dst), + (ins _SrcRC.MemOp:$src), + !strconcat(asm,"\t{$src, $dst|$dst, $src}"), + []>, EVEX, VEX_LIG; + + } // isCodeGenOnly = 1, hasSideEffects = 0 +} //HasAVX512 +} + + +defm VCVTTSS2SIZ: avx512_cvt_s_all<0x2C, "cvttss2si", f32x_info, i32x_info, + fp_to_sint,X86cvttss2IntRnd>, + XS, EVEX_CD8<32, CD8VT1>; +defm VCVTTSS2SI64Z: avx512_cvt_s_all<0x2C, "cvttss2si", f32x_info, i64x_info, + fp_to_sint,X86cvttss2IntRnd>, + VEX_W, XS, EVEX_CD8<32, CD8VT1>; +defm VCVTTSD2SIZ: avx512_cvt_s_all<0x2C, "cvttsd2si", f64x_info, i32x_info, + fp_to_sint,X86cvttsd2IntRnd>, + XD, EVEX_CD8<64, CD8VT1>; +defm VCVTTSD2SI64Z: avx512_cvt_s_all<0x2C, "cvttsd2si", f64x_info, i64x_info, + fp_to_sint,X86cvttsd2IntRnd>, + VEX_W, XD, EVEX_CD8<64, CD8VT1>; + +defm VCVTTSS2USIZ: avx512_cvt_s_all<0x78, "cvttss2usi", f32x_info, i32x_info, + fp_to_uint,X86cvttss2UIntRnd>, + XS, EVEX_CD8<32, CD8VT1>; +defm VCVTTSS2USI64Z: avx512_cvt_s_all<0x78, "cvttss2usi", f32x_info, i64x_info, + fp_to_uint,X86cvttss2UIntRnd>, + XS,VEX_W, EVEX_CD8<32, CD8VT1>; +defm VCVTTSD2USIZ: avx512_cvt_s_all<0x78, "cvttsd2usi", f64x_info, i32x_info, + fp_to_uint,X86cvttsd2UIntRnd>, + XD, EVEX_CD8<64, CD8VT1>; +defm VCVTTSD2USI64Z: avx512_cvt_s_all<0x78, "cvttsd2usi", f64x_info, i64x_info, + fp_to_uint,X86cvttsd2UIntRnd>, + XD, VEX_W, EVEX_CD8<64, CD8VT1>; +let Predicates = [HasAVX512] in { + def : Pat<(i32 (int_x86_sse_cvttss2si (v4f32 VR128X:$src))), + (VCVTTSS2SIZrr_Int (COPY_TO_REGCLASS VR128X:$src, FR32X))>; + def : Pat<(i64 (int_x86_sse_cvttss2si64 (v4f32 VR128X:$src))), + (VCVTTSS2SI64Zrr_Int (COPY_TO_REGCLASS VR128X:$src, FR32X))>; + def : Pat<(i32 (int_x86_sse2_cvttsd2si (v2f64 VR128X:$src))), + (VCVTTSD2SIZrr_Int (COPY_TO_REGCLASS VR128X:$src, FR64X))>; + def : Pat<(i64 (int_x86_sse2_cvttsd2si64 (v2f64 VR128X:$src))), + (VCVTTSD2SI64Zrr_Int (COPY_TO_REGCLASS VR128X:$src, FR64X))>; + } // HasAVX512 //===----------------------------------------------------------------------===// // AVX-512 Convert form float to double and back //===----------------------------------------------------------------------===// -let hasSideEffects = 0 in { -def VCVTSS2SDZrr : AVX512XSI<0x5A, MRMSrcReg, (outs FR64X:$dst), - (ins FR32X:$src1, FR32X:$src2), - "vcvtss2sd\t{$src2, $src1, $dst|$dst, $src1, $src2}", - []>, EVEX_4V, VEX_LIG, Sched<[WriteCvtF2F]>; -let mayLoad = 1 in -def VCVTSS2SDZrm : AVX512XSI<0x5A, MRMSrcMem, (outs FR64X:$dst), - (ins FR32X:$src1, f32mem:$src2), - "vcvtss2sd\t{$src2, $src1, $dst|$dst, $src1, $src2}", - []>, EVEX_4V, VEX_LIG, Sched<[WriteCvtF2FLd, ReadAfterLd]>, - EVEX_CD8<32, CD8VT1>; - -// Convert scalar double to scalar single -def VCVTSD2SSZrr : AVX512XDI<0x5A, MRMSrcReg, (outs FR32X:$dst), - (ins FR64X:$src1, FR64X:$src2), - "vcvtsd2ss\t{$src2, $src1, $dst|$dst, $src1, $src2}", - []>, EVEX_4V, VEX_LIG, VEX_W, Sched<[WriteCvtF2F]>; -let mayLoad = 1 in -def VCVTSD2SSZrm : AVX512XDI<0x5A, MRMSrcMem, (outs FR32X:$dst), - (ins FR64X:$src1, f64mem:$src2), - "vcvtsd2ss\t{$src2, $src1, $dst|$dst, $src1, $src2}", - []>, EVEX_4V, VEX_LIG, VEX_W, - Sched<[WriteCvtF2FLd, ReadAfterLd]>, EVEX_CD8<64, CD8VT1>; -} - -def : Pat<(f64 (fextend FR32X:$src)), (VCVTSS2SDZrr FR32X:$src, FR32X:$src)>, - Requires<[HasAVX512]>; -def : Pat<(fextend (loadf32 addr:$src)), - (VCVTSS2SDZrm (f32 (IMPLICIT_DEF)), addr:$src)>, Requires<[HasAVX512]>; - -def : Pat<(extloadf32 addr:$src), - (VCVTSS2SDZrm (f32 (IMPLICIT_DEF)), addr:$src)>, +multiclass avx512_cvt_fp_scalar<bits<8> opc, string OpcodeStr, X86VectorVTInfo _, + X86VectorVTInfo _Src, SDNode OpNode> { + defm rr : AVX512_maskable_scalar<opc, MRMSrcReg, _, (outs _.RC:$dst), + (ins _Src.RC:$src1, _Src.RC:$src2), OpcodeStr, + "$src2, $src1", "$src1, $src2", + (_.VT (OpNode (_Src.VT _Src.RC:$src1), + (_Src.VT _Src.RC:$src2)))>, + EVEX_4V, VEX_LIG, Sched<[WriteCvtF2F]>; + defm rm : AVX512_maskable_scalar<opc, MRMSrcMem, _, (outs _.RC:$dst), + (ins _Src.RC:$src1, _Src.MemOp:$src2), OpcodeStr, + "$src2, $src1", "$src1, $src2", + (_.VT (OpNode (_Src.VT _Src.RC:$src1), + (_Src.VT (scalar_to_vector + (_Src.ScalarLdFrag addr:$src2)))))>, + EVEX_4V, VEX_LIG, Sched<[WriteCvtF2FLd, ReadAfterLd]>; +} + +// Scalar Coversion with SAE - suppress all exceptions +multiclass avx512_cvt_fp_sae_scalar<bits<8> opc, string OpcodeStr, X86VectorVTInfo _, + X86VectorVTInfo _Src, SDNode OpNodeRnd> { + defm rrb : AVX512_maskable_scalar<opc, MRMSrcReg, _, (outs _.RC:$dst), + (ins _Src.RC:$src1, _Src.RC:$src2), OpcodeStr, + "{sae}, $src2, $src1", "$src1, $src2, {sae}", + (_.VT (OpNodeRnd (_Src.VT _Src.RC:$src1), + (_Src.VT _Src.RC:$src2), + (i32 FROUND_NO_EXC)))>, + EVEX_4V, VEX_LIG, EVEX_B; +} + +// Scalar Conversion with rounding control (RC) +multiclass avx512_cvt_fp_rc_scalar<bits<8> opc, string OpcodeStr, X86VectorVTInfo _, + X86VectorVTInfo _Src, SDNode OpNodeRnd> { + defm rrb : AVX512_maskable_scalar<opc, MRMSrcReg, _, (outs _.RC:$dst), + (ins _Src.RC:$src1, _Src.RC:$src2, AVX512RC:$rc), OpcodeStr, + "$rc, $src2, $src1", "$src1, $src2, $rc", + (_.VT (OpNodeRnd (_Src.VT _Src.RC:$src1), + (_Src.VT _Src.RC:$src2), (i32 imm:$rc)))>, + EVEX_4V, VEX_LIG, Sched<[WriteCvtF2FLd, ReadAfterLd]>, + EVEX_B, EVEX_RC; +} +multiclass avx512_cvt_fp_scalar_sd2ss<bits<8> opc, string OpcodeStr, SDNode OpNode, + SDNode OpNodeRnd, X86VectorVTInfo _src, + X86VectorVTInfo _dst> { + let Predicates = [HasAVX512] in { + defm Z : avx512_cvt_fp_scalar<opc, OpcodeStr, _dst, _src, OpNode>, + avx512_cvt_fp_rc_scalar<opc, OpcodeStr, _dst, _src, + OpNodeRnd>, VEX_W, EVEX_CD8<64, CD8VT1>, + EVEX_V512, XD; + } +} + +multiclass avx512_cvt_fp_scalar_ss2sd<bits<8> opc, string OpcodeStr, SDNode OpNode, + SDNode OpNodeRnd, X86VectorVTInfo _src, + X86VectorVTInfo _dst> { + let Predicates = [HasAVX512] in { + defm Z : avx512_cvt_fp_scalar<opc, OpcodeStr, _dst, _src, OpNode>, + avx512_cvt_fp_sae_scalar<opc, OpcodeStr, _dst, _src, OpNodeRnd>, + EVEX_CD8<32, CD8VT1>, XS, EVEX_V512; + } +} +defm VCVTSD2SS : avx512_cvt_fp_scalar_sd2ss<0x5A, "vcvtsd2ss", X86fround, + X86froundRnd, f64x_info, f32x_info>; +defm VCVTSS2SD : avx512_cvt_fp_scalar_ss2sd<0x5A, "vcvtss2sd", X86fpext, + X86fpextRnd,f32x_info, f64x_info >; + +def : Pat<(f64 (fextend FR32X:$src)), + (COPY_TO_REGCLASS (VCVTSS2SDZrr (COPY_TO_REGCLASS FR32X:$src, VR128X), + (COPY_TO_REGCLASS FR32X:$src, VR128X)), VR128X)>, + Requires<[HasAVX512]>; +def : Pat<(f64 (fextend (loadf32 addr:$src))), + (COPY_TO_REGCLASS (VCVTSS2SDZrm (v4f32 (IMPLICIT_DEF)), addr:$src), VR128X)>, + Requires<[HasAVX512]>; + +def : Pat<(f64 (extloadf32 addr:$src)), + (COPY_TO_REGCLASS (VCVTSS2SDZrm (v4f32 (IMPLICIT_DEF)), addr:$src), VR128X)>, Requires<[HasAVX512, OptForSize]>; -def : Pat<(extloadf32 addr:$src), - (VCVTSS2SDZrr (f32 (IMPLICIT_DEF)), (VMOVSSZrm addr:$src))>, - Requires<[HasAVX512, OptForSpeed]>; +def : Pat<(f64 (extloadf32 addr:$src)), + (COPY_TO_REGCLASS (VCVTSS2SDZrr (v4f32 (IMPLICIT_DEF)), + (COPY_TO_REGCLASS (VMOVSSZrm addr:$src), VR128X)), VR128X)>, + Requires<[HasAVX512, OptForSpeed]>; -def : Pat<(f32 (fround FR64X:$src)), (VCVTSD2SSZrr FR64X:$src, FR64X:$src)>, +def : Pat<(f32 (fround FR64X:$src)), + (COPY_TO_REGCLASS (VCVTSD2SSZrr (COPY_TO_REGCLASS FR64X:$src, VR128X), + (COPY_TO_REGCLASS FR64X:$src, VR128X)), VR128X)>, Requires<[HasAVX512]>; - //===----------------------------------------------------------------------===// // AVX-512 Vector convert from signed/unsigned integer to float/double // and from float/double to signed/unsigned integer |
