← Back to branch summary

~pipelight/pipelight/llvm-3.4

« back to all changes in this revision

Viewing changes to clang/lib/CodeGen/CGBuiltin.cpp

  • Committer: Package Import Robot
  • Author(s): Sylvestre Ledru
  • Date: 2013-12-03 10:25:59 UTC
  • mfrom: (0.10.1) (0.9.1) (0.8.1) (0.7.1) (0.6.1) (0.2.2)
  • Revision ID: package-import@ubuntu.com-20131203102559-chcuo5tqergytnct
Tags: 1:3.4~+rc2-1
* New testing upstream release
* 0047-version-name.diff ocamldoc.diff removed (applied upstream)
* r600 is now compiled by default (remove the configure arg)

Show diffs side-by-side

added added

removed removed

Lines of Context:
clang/lib/CodeGen/CGBuiltin.cpp
1759
1759
  bool ExtendEle = false;
1760
1760
  bool OverloadInt = false;
1761
1761
  bool OverloadCmpInt = false;
 
1762
  bool OverloadCvtInt = false;
1762
1763
  bool OverloadWideInt = false;
1763
1764
  bool OverloadNarrowInt = false;
1764
1765
  const char *s = NULL;
1772
1773
  // argument that specifies the vector type, need to handle each case.
1773
1774
  switch (BuiltinID) {
1774
1775
  default: break;
 
1776
  case AArch64::BI__builtin_neon_vdups_lane_f32:
 
1777
  case AArch64::BI__builtin_neon_vdupd_lane_f64:
 
1778
  case AArch64::BI__builtin_neon_vdups_laneq_f32:
 
1779
  case AArch64::BI__builtin_neon_vdupd_laneq_f64: {
 
1780
    return CGF.Builder.CreateExtractElement(Ops[0], Ops[1], "vdup_lane");
 
1781
  }
 
1782
  case AArch64::BI__builtin_neon_vdupb_lane_i8:
 
1783
  case AArch64::BI__builtin_neon_vduph_lane_i16:
 
1784
  case AArch64::BI__builtin_neon_vdups_lane_i32:
 
1785
  case AArch64::BI__builtin_neon_vdupd_lane_i64:
 
1786
  case AArch64::BI__builtin_neon_vdupb_laneq_i8:
 
1787
  case AArch64::BI__builtin_neon_vduph_laneq_i16:
 
1788
  case AArch64::BI__builtin_neon_vdups_laneq_i32:
 
1789
  case AArch64::BI__builtin_neon_vdupd_laneq_i64: {
 
1790
    // The backend treats Neon scalar types as v1ix types
 
1791
    // So we want to dup lane from any vector to v1ix vector
 
1792
    // with shufflevector
 
1793
    s = "vdup_lane";
 
1794
    Value* SV = llvm::ConstantVector::getSplat(1, cast<ConstantInt>(Ops[1]));
 
1795
    Value *Result = CGF.Builder.CreateShuffleVector(Ops[0], Ops[0], SV, s);
 
1796
    llvm::Type *Ty = CGF.ConvertType(E->getCallReturnType());
 
1797
    // AArch64 intrinsic one-element vector type cast to
 
1798
    // scalar type expected by the builtin
 
1799
    return CGF.Builder.CreateBitCast(Result, Ty, s);
 
1800
  }
 
1801
  case AArch64::BI__builtin_neon_vqdmlalh_lane_s16 :
 
1802
  case AArch64::BI__builtin_neon_vqdmlalh_laneq_s16 :
 
1803
  case AArch64::BI__builtin_neon_vqdmlals_lane_s32 :
 
1804
  case AArch64::BI__builtin_neon_vqdmlals_laneq_s32 :
 
1805
  case AArch64::BI__builtin_neon_vqdmlslh_lane_s16 :
 
1806
  case AArch64::BI__builtin_neon_vqdmlslh_laneq_s16 :
 
1807
  case AArch64::BI__builtin_neon_vqdmlsls_lane_s32 :
 
1808
  case AArch64::BI__builtin_neon_vqdmlsls_laneq_s32 : {
 
1809
    Int = Intrinsic::arm_neon_vqadds;
 
1810
    if (BuiltinID == AArch64::BI__builtin_neon_vqdmlslh_lane_s16 ||
 
1811
        BuiltinID == AArch64::BI__builtin_neon_vqdmlslh_laneq_s16 ||
 
1812
        BuiltinID == AArch64::BI__builtin_neon_vqdmlsls_lane_s32 ||
 
1813
        BuiltinID == AArch64::BI__builtin_neon_vqdmlsls_laneq_s32) {
 
1814
      Int = Intrinsic::arm_neon_vqsubs;
 
1815
    }
 
1816
    // create vqdmull call with b * c[i]
 
1817
    llvm::Type *Ty = CGF.ConvertType(E->getArg(1)->getType());
 
1818
    llvm::VectorType *OpVTy = llvm::VectorType::get(Ty, 1);
 
1819
    Ty = CGF.ConvertType(E->getArg(0)->getType());
 
1820
    llvm::VectorType *ResVTy = llvm::VectorType::get(Ty, 1);
 
1821
    Value *F = CGF.CGM.getIntrinsic(Intrinsic::arm_neon_vqdmull, ResVTy);
 
1822
    Value *V = UndefValue::get(OpVTy);
 
1823
    llvm::Constant *CI = ConstantInt::get(CGF.Int32Ty, 0);
 
1824
    SmallVector<Value *, 2> MulOps;
 
1825
    MulOps.push_back(Ops[1]);
 
1826
    MulOps.push_back(Ops[2]);
 
1827
    MulOps[0] = CGF.Builder.CreateInsertElement(V, MulOps[0], CI);
 
1828
    MulOps[1] = CGF.Builder.CreateExtractElement(MulOps[1], Ops[3], "extract");
 
1829
    MulOps[1] = CGF.Builder.CreateInsertElement(V, MulOps[1], CI);
 
1830
    Value *MulRes = CGF.Builder.CreateCall2(F, MulOps[0], MulOps[1]);
 
1831
    // create vqadds call with a +/- vqdmull result
 
1832
    F = CGF.CGM.getIntrinsic(Int, ResVTy);
 
1833
    SmallVector<Value *, 2> AddOps;
 
1834
    AddOps.push_back(Ops[0]);
 
1835
    AddOps.push_back(MulRes);
 
1836
    V = UndefValue::get(ResVTy);
 
1837
    AddOps[0] = CGF.Builder.CreateInsertElement(V, AddOps[0], CI);
 
1838
    Value *AddRes = CGF.Builder.CreateCall2(F, AddOps[0], AddOps[1]);
 
1839
    return CGF.Builder.CreateBitCast(AddRes, Ty);
 
1840
  }
1775
1841
  case AArch64::BI__builtin_neon_vfmas_lane_f32:
1776
1842
  case AArch64::BI__builtin_neon_vfmas_laneq_f32:
1777
1843
  case AArch64::BI__builtin_neon_vfmad_lane_f64:
1778
1844
  case AArch64::BI__builtin_neon_vfmad_laneq_f64: {
1779
1845
    llvm::Type *Ty = CGF.ConvertType(E->getCallReturnType());
1780
1846
    Value *F = CGF.CGM.getIntrinsic(Intrinsic::fma, Ty);
1781
 
    // extract lane acc += x * v[i]
1782
1847
    Ops[2] = CGF.Builder.CreateExtractElement(Ops[2], Ops[3], "extract");
1783
1848
    return CGF.Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]);
1784
1849
  }
1857
1922
  case AArch64::BI__builtin_neon_vqaddh_s16:
1858
1923
  case AArch64::BI__builtin_neon_vqadds_s32:
1859
1924
  case AArch64::BI__builtin_neon_vqaddd_s64:
1860
 
    Int = Intrinsic::aarch64_neon_vqadds;
 
1925
    Int = Intrinsic::arm_neon_vqadds;
1861
1926
    s = "vqadds"; OverloadInt = true; break;
1862
1927
  case AArch64::BI__builtin_neon_vqaddb_u8:
1863
1928
  case AArch64::BI__builtin_neon_vqaddh_u16:
1864
1929
  case AArch64::BI__builtin_neon_vqadds_u32:
1865
1930
  case AArch64::BI__builtin_neon_vqaddd_u64:
1866
 
    Int = Intrinsic::aarch64_neon_vqaddu;
 
1931
    Int = Intrinsic::arm_neon_vqaddu;
1867
1932
    s = "vqaddu"; OverloadInt = true; break;
1868
1933
  // Scalar Saturating Sub
1869
1934
  case AArch64::BI__builtin_neon_vqsubb_s8:
1870
1935
  case AArch64::BI__builtin_neon_vqsubh_s16:
1871
1936
  case AArch64::BI__builtin_neon_vqsubs_s32:
1872
1937
  case AArch64::BI__builtin_neon_vqsubd_s64:
1873
 
    Int = Intrinsic::aarch64_neon_vqsubs;
 
1938
    Int = Intrinsic::arm_neon_vqsubs;
1874
1939
    s = "vqsubs"; OverloadInt = true; break;
1875
1940
  case AArch64::BI__builtin_neon_vqsubb_u8:
1876
1941
  case AArch64::BI__builtin_neon_vqsubh_u16:
1877
1942
  case AArch64::BI__builtin_neon_vqsubs_u32:
1878
1943
  case AArch64::BI__builtin_neon_vqsubd_u64:
1879
 
    Int = Intrinsic::aarch64_neon_vqsubu;
 
1944
    Int = Intrinsic::arm_neon_vqsubu;
1880
1945
    s = "vqsubu"; OverloadInt = true; break;
1881
1946
  // Scalar Shift Left
1882
1947
  case AArch64::BI__builtin_neon_vshld_s64:
2057
2122
  case AArch64::BI__builtin_neon_vcvtd_f64_u64:
2058
2123
    Int = Intrinsic::aarch64_neon_vcvtf64_u64,
2059
2124
    s = "vcvtf"; OverloadInt = false; break;
 
2125
  // Scalar Floating-point Converts
 
2126
  case AArch64::BI__builtin_neon_vcvtxd_f32_f64:
 
2127
    Int = Intrinsic::aarch64_neon_fcvtxn;
 
2128
    s = "vcvtxn"; OverloadCvtInt = true; break;
 
2129
  case AArch64::BI__builtin_neon_vcvtas_s32_f32:
 
2130
  case AArch64::BI__builtin_neon_vcvtad_s64_f64:
 
2131
    Int = Intrinsic::aarch64_neon_fcvtas;
 
2132
    s = "vcvtas"; OverloadCvtInt = true; break;
 
2133
  case AArch64::BI__builtin_neon_vcvtas_u32_f32:
 
2134
  case AArch64::BI__builtin_neon_vcvtad_u64_f64:
 
2135
    Int = Intrinsic::aarch64_neon_fcvtau;
 
2136
    s = "vcvtau"; OverloadCvtInt = true; break;
 
2137
  case AArch64::BI__builtin_neon_vcvtms_s32_f32:
 
2138
  case AArch64::BI__builtin_neon_vcvtmd_s64_f64:
 
2139
    Int = Intrinsic::aarch64_neon_fcvtms;
 
2140
    s = "vcvtms"; OverloadCvtInt = true; break;
 
2141
  case AArch64::BI__builtin_neon_vcvtms_u32_f32:
 
2142
  case AArch64::BI__builtin_neon_vcvtmd_u64_f64:
 
2143
    Int = Intrinsic::aarch64_neon_fcvtmu;
 
2144
    s = "vcvtmu"; OverloadCvtInt = true; break;
 
2145
  case AArch64::BI__builtin_neon_vcvtns_s32_f32:
 
2146
  case AArch64::BI__builtin_neon_vcvtnd_s64_f64:
 
2147
    Int = Intrinsic::aarch64_neon_fcvtns;
 
2148
    s = "vcvtns"; OverloadCvtInt = true; break;
 
2149
  case AArch64::BI__builtin_neon_vcvtns_u32_f32:
 
2150
  case AArch64::BI__builtin_neon_vcvtnd_u64_f64:
 
2151
    Int = Intrinsic::aarch64_neon_fcvtnu;
 
2152
    s = "vcvtnu"; OverloadCvtInt = true; break;
 
2153
  case AArch64::BI__builtin_neon_vcvtps_s32_f32:
 
2154
  case AArch64::BI__builtin_neon_vcvtpd_s64_f64:
 
2155
    Int = Intrinsic::aarch64_neon_fcvtps;
 
2156
    s = "vcvtps"; OverloadCvtInt = true; break;
 
2157
  case AArch64::BI__builtin_neon_vcvtps_u32_f32:
 
2158
  case AArch64::BI__builtin_neon_vcvtpd_u64_f64:
 
2159
    Int = Intrinsic::aarch64_neon_fcvtpu;
 
2160
    s = "vcvtpu"; OverloadCvtInt = true; break;
 
2161
  case AArch64::BI__builtin_neon_vcvts_s32_f32:
 
2162
  case AArch64::BI__builtin_neon_vcvtd_s64_f64:
 
2163
    Int = Intrinsic::aarch64_neon_fcvtzs;
 
2164
    s = "vcvtzs"; OverloadCvtInt = true; break;
 
2165
  case AArch64::BI__builtin_neon_vcvts_u32_f32:
 
2166
  case AArch64::BI__builtin_neon_vcvtd_u64_f64:
 
2167
    Int = Intrinsic::aarch64_neon_fcvtzu;
 
2168
    s = "vcvtzu"; OverloadCvtInt = true; break;
2060
2169
  // Scalar Floating-point Reciprocal Estimate
2061
2170
  case AArch64::BI__builtin_neon_vrecpes_f32:
2062
2171
  case AArch64::BI__builtin_neon_vrecped_f64:
2225
2334
  case AArch64::BI__builtin_neon_vabsd_s64:
2226
2335
    Int = Intrinsic::aarch64_neon_vabs;
2227
2336
    s = "vabs"; OverloadInt = false; break;
 
2337
  // Scalar Absolute Difference
 
2338
  case AArch64::BI__builtin_neon_vabds_f32:
 
2339
  case AArch64::BI__builtin_neon_vabdd_f64:
 
2340
    Int = Intrinsic::aarch64_neon_vabd;
 
2341
    s = "vabd"; OverloadInt = true; break;
2228
2342
  // Scalar Signed Saturating Absolute Value
2229
2343
  case AArch64::BI__builtin_neon_vqabsb_s8:
2230
2344
  case AArch64::BI__builtin_neon_vqabsh_s16:
2270
2384
  // Signed Saturating Doubling Multiply Long
2271
2385
  case AArch64::BI__builtin_neon_vqdmullh_s16:
2272
2386
  case AArch64::BI__builtin_neon_vqdmulls_s32:
2273
 
    Int = Intrinsic::aarch64_neon_vqdmull;
 
2387
    Int = Intrinsic::arm_neon_vqdmull;
2274
2388
    s = "vqdmull"; OverloadWideInt = true; break;
2275
2389
  // Scalar Signed Saturating Extract Unsigned Narrow
2276
2390
  case AArch64::BI__builtin_neon_vqmovunh_s16:
2476
2590
    Tys.push_back(VTy);
2477
2591
 
2478
2592
    F = CGF.CGM.getIntrinsic(Int, Tys);
 
2593
  } else if (OverloadCvtInt) {
 
2594
    // Determine the types of this overloaded AArch64 intrinsic
 
2595
    SmallVector<llvm::Type *, 2> Tys;
 
2596
    const Expr *Arg = E->getArg(E->getNumArgs()-1);
 
2597
    llvm::Type *Ty = CGF.ConvertType(E->getCallReturnType());
 
2598
    llvm::VectorType *VTy = llvm::VectorType::get(Ty, 1);
 
2599
    Tys.push_back(VTy);
 
2600
    Ty = CGF.ConvertType(Arg->getType());
 
2601
    VTy = llvm::VectorType::get(Ty, 1);
 
2602
    Tys.push_back(VTy);
 
2603
 
 
2604
    F = CGF.CGM.getIntrinsic(Int, Tys);
2479
2605
  } else
2480
2606
    F = CGF.CGM.getIntrinsic(Int);
2481
2607
 
3543
3669
    return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt")
3544
3670
                : Builder.CreateSIToFP(Ops[0], Ty, "vcvt");
3545
3671
  }
3546
 
    return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vqmovun_v, E);
3547
3672
  case AArch64::BI__builtin_neon_vrndn_v:
3548
3673
  case AArch64::BI__builtin_neon_vrndnq_v: {
3549
3674
    Int = Intrinsic::aarch64_neon_frintn;