xref: /third_party/vixl/test/aarch64/test-assembler-neon-aarch64.cc

315   __ Ld1(v5.V4H(), v6.V4H(), v7.V4H(), MemOperand(x17));
369   __ Ld1(v5.V4H(), v6.V4H(), v7.V4H(), MemOperand(x19, 24, PostIndex));
596   __ Ld2(v6.V4H(), v7.V4H(), MemOperand(x17));
633   __ Ld2(v5.V4H(), v6.V4H(), MemOperand(x19, 16, PostIndex));
938   __ Ld2r(v4.V4H(), v5.V4H(), MemOperand(x17));
984   __ Ld2r(v4.V4H(), v5.V4H(), MemOperand(x17, x18, PostIndex));
1028   __ Ld3(v8.V4H(), v9.V4H(), v10.V4H(), MemOperand(x17));
1070   __ Ld3(v8.V4H(), v9.V4H(), v10.V4H(), MemOperand(x19, 24, PostIndex));
1409   __ Ld3r(v6.V4H(), v7.V4H(), v8.V4H(), MemOperand(x17));
1464   __ Ld3r(v6.V4H(), v7.V4H(), v8.V4H(), MemOperand(x17, x18, PostIndex));
1514   __ Ld4(v10.V4H(), v11.V4H(), v12.V4H(), v13.V4H(), MemOperand(x17));
1568   __ Ld4(v10.V4H(),
1569          v11.V4H(),
1570          v12.V4H(),
1571          v13.V4H(),
2014   __ Ld4r(v8.V4H(), v9.V4H(), v10.V4H(), v11.V4H(), MemOperand(x17));
2086   __ Ld4r(v8.V4H(),
2087           v9.V4H(),
2088           v10.V4H(),
2089           v11.V4H(),
2615   __ Ld1r(v2.V4H(), MemOperand(x17));
2657   __ Ld1r(v2.V4H(), MemOperand(x17, x18, PostIndex));
2702   __ St1(v0.V4H(), v1.V4H(), v2.V4H(), MemOperand(x17));
2763   __ St1(v0.V4H(), v1.V4H(), v2.V4H(), MemOperand(x17, 24, PostIndex));
2931   __ St2(v0.V4H(), v1.V4H(), MemOperand(x18));
2971   __ St2(v0.V4H(), v1.V4H(), MemOperand(x18, 16, PostIndex));
3094   __ St3(v0.V4H(), v1.V4H(), v2.V4H(), MemOperand(x18));
3132   __ St3(v0.V4H(), v1.V4H(), v2.V4H(), MemOperand(x18, 24, PostIndex));
3268   __ St4(v0.V4H(), v1.V4H(), v2.V4H(), v3.V4H(), MemOperand(x18));
3315   __ St4(v0.V4H(),
3316          v1.V4H(),
3317          v2.V4H(),
3318          v3.V4H(),
3650   __ Fcvtl(v20.V4S(), v1.V4H());
3654   __ Fcvtl(v22.V4S(), v22.V4H());
3680   __ Fmov(v0.V4H(), 24.0);
3681   __ Fmov(v1.V4H(), 1024.0);
3686   __ Fmov(v6.V4H(), RawbitsToFloat16(0x7c2f));
3689   __ Fadd(v8.V4H(), v1.V4H(), v0.V4H());
3691   __ Fadd(v10.V4H(), v4.V4H(), v3.V4H());
3693   __ Fadd(v11.V4H(), v6.V4H(), v1.V4H());
3694   __ Fadd(v12.V4H(), v7.V4H(), v7.V4H());
3722   __ Fmov(v0.V4H(), 24.0);
3723   __ Fmov(v1.V4H(), 1024.0);
3726   __ Fmov(v4.V4H(), kFP16PositiveInfinity);
3727   __ Fmov(v5.V4H(), kFP16NegativeInfinity);
3728   __ Fmov(v6.V4H(), RawbitsToFloat16(0x7c22));
3731   __ Fsub(v0.V4H(), v1.V4H(), v0.V4H());
3733   __ Fsub(v9.V4H(), v4.V4H(), v3.V4H());
3734   __ Fsub(v10.V4H(), v0.V4H(), v1.V4H());
3736   __ Fsub(v11.V4H(), v6.V4H(), v2.V4H());
3737   __ Fsub(v12.V4H(), v7.V4H(), v7.V4H());
3765   __ Fmov(v0.V4H(), 24.0);
3766   __ Fmov(v1.V4H(), -2.0);
3769   __ Fmov(v4.V4H(), kFP16PositiveInfinity);
3770   __ Fmov(v5.V4H(), kFP16NegativeInfinity);
3772   __ Fmul(v6.V4H(), v1.V4H(), v0.V4H());
3774   __ Fmul(v8.V4H(), v4.V4H(), v3.V4H());
3775   __ Fmul(v9.V4H(), v0.V4H(), v1.V4H());
3776   __ Fmul(v10.V4H(), v5.V4H(), v0.V4H());
3796   __ Fmov(v0.V4H(), 24.0);
3797   __ Fmov(v1.V4H(), -2.0);
3800   __ Fmov(v4.V4H(), kFP16PositiveInfinity);
3801   __ Fmov(v5.V4H(), kFP16NegativeInfinity);
3803   __ Fdiv(v6.V4H(), v0.V4H(), v1.V4H());
3805   __ Fdiv(v8.V4H(), v4.V4H(), v3.V4H());
3806   __ Fdiv(v9.V4H(), v1.V4H(), v0.V4H());
3807   __ Fdiv(v10.V4H(), v5.V4H(), v0.V4H());
3832   __ Fcvtl(v16.V4S(), v0.V4H());
3834   __ Fcvtl(v18.V4S(), v1.V4H());
3877   __ Fcvtn(v16.V4H(), v0.V4S());
3879   __ Fcvtn(v17.V4H(), v2.V4S());
3907   __ Fcvtn(v17.V4H(), v1.V4S());
3909   __ Fcvtn(v1.V4H(), v1.V4S());
3988   __ Sqdmulh(v16.V4H(), v0.V4H(), v1.V4H());
3993   __ Sqrdmulh(v20.V4H(), v0.V4H(), v1.V4H());
4023   __ Sqdmulh(v16.V4H(), v0.V4H(), v1.H(), 1);
4028   __ Sqrdmulh(v20.V4H(), v0.V4H(), v1.H(), 1);
4063   __ Sqrdmlah(v16.V4H(), v0.V4H(), v1.V4H());
4094   __ Sqrdmlah(v16.V4H(), v0.V4H(), v1.H(), 1);
4125   __ Sqrdmlsh(v16.V4H(), v0.V4H(), v1.V4H());
4156   __ Sqrdmlsh(v16.V4H(), v0.V4H(), v1.H(), 1);
4243   __ Saddlp(v17.V4H(), v0.V8B());
4246   __ Saddlp(v19.V2S(), v0.V4H());
4272   __ Uaddlp(v17.V4H(), v0.V8B());
4275   __ Uaddlp(v19.V2S(), v0.V4H());
4307   __ Sadalp(v17.V4H(), v0.V8B());
4312   __ Sadalp(v19.V2S(), v1.V4H());
4346   __ Uadalp(v17.V4H(), v0.V8B());
4351   __ Uadalp(v19.V2S(), v1.V4H());
4432   __ Mul(v16.V4H(), v0.V4H(), v1.H(), 0);
4439   __ Mla(v20.V4H(), v0.V4H(), v1.H(), 0);
4449   __ Mls(v24.V4H(), v0.V4H(), v1.H(), 0);
4488   __ Smull(v16.V4S(), v0.V4H(), v1.H(), 7);
4490   __ Umull(v18.V4S(), v0.V4H(), v1.H(), 7);
4498   __ Smlal(v20.V4S(), v0.V4H(), v1.H(), 7);
4500   __ Umlal(v22.V4S(), v0.V4H(), v1.H(), 7);
4508   __ Smlsl(v24.V4S(), v0.V4H(), v1.H(), 7);
4510   __ Umlsl(v26.V4S(), v0.V4H(), v1.H(), 7);
4544   __ Sqdmull(v16.V4S(), v0.V4H(), v1.H(), 7);
4552   __ Sqdmlal(v20.V4S(), v0.V4H(), v1.H(), 7);
4560   __ Sqdmlsl(v24.V4S(), v0.V4H(), v1.H(), 7);
4623   __ Sqdmull(v16.V4S(), v0.V4H(), v1.V4H());
4661   __ Sqdmlal(v16.V4S(), v0.V4H(), v1.V4H());
4699   __ Sqdmlsl(v16.V4S(), v0.V4H(), v1.V4H());
4874   __ Fcmeq(v8.V4H(), v0.V4H(), v0.V4H());
4875   __ Fcmeq(v9.V4H(), v1.V4H(), v0.V4H());
4876   __ Fcmeq(v10.V4H(), v2.V4H(), v0.V4H());
4877   __ Fcmeq(v11.V4H(), v3.V4H(), v0.V4H());
4940   __ Fcmge(v8.V4H(), v0.V4H(), v0.V4H());
4941   __ Fcmge(v9.V4H(), v1.V4H(), v0.V4H());
4942   __ Fcmge(v10.V4H(), v2.V4H(), v0.V4H());
4943   __ Fcmge(v11.V4H(), v3.V4H(), v0.V4H());
5006   __ Fcmgt(v8.V4H(), v0.V4H(), v0.V4H());
5007   __ Fcmgt(v9.V4H(), v1.V4H(), v0.V4H());
5008   __ Fcmgt(v10.V4H(), v2.V4H(), v0.V4H());
5009   __ Fcmgt(v11.V4H(), v3.V4H(), v0.V4H());
5072   __ Facge(v8.V4H(), v0.V4H(), v0.V4H());
5073   __ Facge(v9.V4H(), v1.V4H(), v0.V4H());
5074   __ Facge(v10.V4H(), v2.V4H(), v0.V4H());
5075   __ Facge(v11.V4H(), v3.V4H(), v0.V4H());
5138   __ Facgt(v8.V4H(), v0.V4H(), v0.V4H());
5139   __ Facgt(v9.V4H(), v1.V4H(), v0.V4H());
5140   __ Facgt(v10.V4H(), v2.V4H(), v0.V4H());
5141   __ Facgt(v11.V4H(), v3.V4H(), v0.V4H());
5390   __ Cmeq(v18.V4H(), v1.V4H(), 0);
5425   __ Cmge(v18.V4H(), v1.V4H(), 0);
5460   __ Cmlt(v18.V4H(), v1.V4H(), 0);
5495   __ Cmle(v18.V4H(), v1.V4H(), 0);
5530   __ Cmgt(v18.V4H(), v1.V4H(), 0);
5568   __ Neg(v18.V4H(), v1.V4H());
5606   __ Sqneg(v18.V4H(), v1.V4H());
5652   __ Abs(v18.V4H(), v1.V4H());
5690   __ Sqabs(v18.V4H(), v1.V4H());
5754   __ Suqadd(v18.V4H(), v3.V4H());
5816   __ Usqadd(v18.V4H(), v3.V4H());
5859   __ Xtn(v17.V4H(), v1.V4S());
5888   __ Sqxtn(v17.V4H(), v1.V4S());
5923   __ Uqxtn(v17.V4H(), v1.V4S());
5958   __ Sqxtun(v17.V4H(), v1.V4S());
6099   __ Mov(v25.V4H(), v0.V4H());
6266   __ Smax(v18.V4H(), v0.V4H(), v1.V4H());
6296   __ Smaxp(v18.V4H(), v0.V4H(), v1.V4H());
6352   __ Addv(h18, v1.V4H());
6381   __ Saddlv(s18, v1.V4H());
6410   __ Uaddlv(s18, v1.V4H());
6439   __ Smaxv(h18, v1.V4H());
6468   __ Sminv(h18, v1.V4H());
6496   __ Umaxv(h18, v1.V4H());
6525   __ Uminv(h18, v1.V4H());
6552   __ Smin(v18.V4H(), v0.V4H(), v1.V4H());
6582   __ Umax(v18.V4H(), v0.V4H(), v1.V4H());
6612   __ Umin(v18.V4H(), v0.V4H(), v1.V4H());
6842   __ Mvn(v25.V4H(), v0.V4H());
6893   __ Cls(v18.V4H(), v1.V4H());
6900   __ Clz(v24.V4H(), v0.V4H());
6945   __ Rev32(v20.V4H(), v0.V4H());
6950   __ Rev64(v24.V4H(), v0.V4H());
7003   __ Sli(v18.V4H(), v1.V4H(), 8);
7048   __ Sri(v18.V4H(), v1.V4H(), 8);
7087   __ Shrn(v17.V4H(), v1.V4S(), 16);
7116   __ Rshrn(v17.V4H(), v1.V4S(), 16);
7145   __ Uqshrn(v17.V4H(), v1.V4S(), 16);
7181   __ Uqrshrn(v17.V4H(), v1.V4S(), 16);
7217   __ Sqshrn(v17.V4H(), v1.V4S(), 16);
7253   __ Sqrshrn(v17.V4H(), v1.V4S(), 16);
7289   __ Sqshrun(v17.V4H(), v1.V4S(), 16);
7325   __ Sqrshrun(v17.V4H(), v1.V4S(), 16);
7365   __ Bic(v16.V4H(), 0x00, 0);
7366   __ Bic(v17.V4H(), 0xff, 8);
7408   __ Movi(v0.V4H(), 0xabab);
7409   __ Movi(v1.V4H(), 0xab00);
7410   __ Movi(v2.V4H(), 0xabff);
7521   __ Movi(v16.V4H(), 0x00, LSL, 0);
7522   __ Movi(v17.V4H(), 0xff, LSL, 8);
7580   __ Mvni(v16.V4H(), 0x00, LSL, 0);
7581   __ Mvni(v17.V4H(), 0xff, LSL, 8);
7646   __ Orr(v16.V4H(), 0x00, 0);
7647   __ Orr(v17.V4H(), 0xff, 8);
7735   __ Fmov(v2.V4H(), RawbitsToFloat16(0xC500U));
7855   __ Dup(v21.V4H(), v1.H(), 7);
7908   __ Dup(v21.V4H(), w0);
8205   __ Uaddl(v3.V4S(), v5.V4H(), v7.V4H());
8206   __ Uaddl(v4.V4S(), v6.V4H(), v7.V4H());
8490   __ Sshll(v18.V4S(), v1.V4H(), 8);
8522   __ Shll(v18.V4S(), v1.V4H(), 16);
8554   __ Ushll(v18.V4S(), v1.V4H(), 8);
8587   __ Sxtl(v18.V4S(), v1.V4H());
8620   __ Uxtl(v18.V4S(), v1.V4H());
8666   __ Ssra(v18.V4H(), v1.V4H(), 8);
8719   __ Srsra(v18.V4H(), v1.V4H(), 8);
8772   __ Usra(v18.V4H(), v1.V4H(), 8);
8825   __ Ursra(v18.V4H(), v1.V4H(), 8);
9355   __ Fmla(v20.V4H(), v0.V4H(), v1.V4H());
9356   __ Fmla(v21.V4H(), v2.V4H(), v3.V4H());
9357   __ Fmla(v22.V4H(), v2.V4H(), v6.V4H());
9358   __ Fmla(v23.V4H(), v3.V4H(), v6.V4H());
9402   __ Fmls(v20.V4H(), v0.V4H(), v1.V4H());
9403   __ Fmls(v21.V4H(), v2.V4H(), v3.V4H());
9404   __ Fmls(v22.V4H(), v2.V4H(), v6.V4H());
9405   __ Fmls(v23.V4H(), v3.V4H(), v6.V4H());
9471   __ Fmlal(v2.V4S(), v30.V4H(), v31.V4H());
9472   __ Fmlal2(v3.V4S(), v30.V4H(), v31.V4H());
9476   __ Fmlsl(v6.V4S(), v30.V4H(), v31.V4H());
9477   __ Fmlsl2(v7.V4S(), v30.V4H(), v31.V4H());
9577   __ Fmlal(v2.V4S(), v30.V4H(), v10.H(), 2);
9578   __ Fmlal2(v3.V4S(), v30.V4H(), v11.H(), 3);
9582   __ Fmlsl(v6.V4S(), v30.V4H(), v14.H(), 6);
9583   __ Fmlsl2(v7.V4S(), v30.V4H(), v15.H(), 7);
9683   __ Fmulx(v11.V4H(), v0.V4H(), v1.V4H());
9684   __ Fmulx(v12.V4H(), v2.V4H(), v4.V4H());
9685   __ Fmulx(v13.V4H(), v2.V4H(), v5.V4H());
9686   __ Fmulx(v14.V4H(), v3.V4H(), v4.V4H());
9687   __ Fmulx(v15.V4H(), v3.V4H(), v5.V4H());
9754   __ Fabd(v11.V4H(), v1.V4H(), v0.V4H());
9755   __ Fabd(v12.V4H(), v2.V4H(), v3.V4H());
9756   __ Fabd(v13.V4H(), v2.V4H(), v5.V4H());
9757   __ Fabd(v14.V4H(), v3.V4H(), v4.V4H());
9758   __ Fabd(v15.V4H(), v3.V4H(), v5.V4H());
9871   __ Frecps(v9.V4H(), v0.V4H(), v2.V4H());
9872   __ Frecps(v10.V4H(), v1.V4H(), v2.V4H());
9873   __ Frecps(v11.V4H(), v0.V4H(), v3.V4H());
9874   __ Frecps(v12.V4H(), v0.V4H(), v4.V4H());
9938   __ Frsqrts(v9.V4H(), v0.V4H(), v2.V4H());
9939   __ Frsqrts(v10.V4H(), v1.V4H(), v2.V4H());
9940   __ Frsqrts(v11.V4H(), v0.V4H(), v3.V4H());
9941   __ Frsqrts(v12.V4H(), v0.V4H(), v4.V4H());
10002   __ Faddp(v6.V4H(), v1.V4H(), v0.V4H());
10003   __ Faddp(v7.V4H(), v3.V4H(), v2.V4H());
10146   __ Fmax(v10.V4H(), v0.V4H(), v1.V4H());
10147   __ Fmax(v11.V4H(), v2.V4H(), v3.V4H());
10148   __ Fmax(v12.V4H(), v4.V4H(), v0.V4H());
10149   __ Fmax(v13.V4H(), v5.V4H(), v1.V4H());
10180   __ Fmaxp(v8.V4H(), v0.V4H(), v1.V4H());
10181   __ Fmaxp(v9.V4H(), v2.V4H(), v3.V4H());
10212   __ Fmaxnm(v10.V4H(), v0.V4H(), v1.V4H());
10213   __ Fmaxnm(v11.V4H(), v2.V4H(), v3.V4H());
10214   __ Fmaxnm(v12.V4H(), v4.V4H(), v0.V4H());
10215   __ Fmaxnm(v13.V4H(), v5.V4H(), v1.V4H());
10246   __ Fmaxnmp(v8.V4H(), v0.V4H(), v1.V4H());
10247   __ Fmaxnmp(v9.V4H(), v2.V4H(), v3.V4H());
10390   __ Fmin(v10.V4H(), v0.V4H(), v1.V4H());
10391   __ Fmin(v11.V4H(), v2.V4H(), v3.V4H());
10392   __ Fmin(v12.V4H(), v4.V4H(), v0.V4H());
10393   __ Fmin(v13.V4H(), v5.V4H(), v1.V4H());
10424   __ Fminp(v8.V4H(), v0.V4H(), v1.V4H());
10425   __ Fminp(v9.V4H(), v2.V4H(), v3.V4H());
10456   __ Fminnm(v10.V4H(), v0.V4H(), v1.V4H());
10457   __ Fminnm(v11.V4H(), v2.V4H(), v3.V4H());
10458   __ Fminnm(v12.V4H(), v4.V4H(), v0.V4H());
10459   __ Fminnm(v13.V4H(), v5.V4H(), v1.V4H());
10490   __ Fminnmp(v8.V4H(), v0.V4H(), v1.V4H());
10491   __ Fminnmp(v9.V4H(), v2.V4H(), v3.V4H());
10583   __ Fmin(v2.V4H(), v0.V4H(), v1.V4H());
10586   __ Fmax(v4.V4H(), v0.V4H(), v1.V4H());
10589   __ Fminnm(v6.V4H(), v0.V4H(), v1.V4H());
10592   __ Fmaxnm(v8.V4H(), v0.V4H(), v1.V4H());

Indexes created Thu Nov 07 10:32:03 CST 2024