[AArch64] Refactor the Neon vector/scalar floating-point convert intrinsics so

that they use float/double rather than the vector equivalents when appropriate.

llvm-svn: 196930

llvm/include/llvm/IR/IntrinsicsAArch64.td

@@ -36,11 +36,11 @@ def int_aarch64_neon_xtn :
 // Vector floating-point convert
 def int_aarch64_neon_frintn : Neon_1Arg_Intrinsic;
 def int_aarch64_neon_fsqrt : Neon_1Arg_Intrinsic;
-def int_aarch64_neon_fcvtxn :
+def int_aarch64_neon_vcvtxn :
   Intrinsic<[llvm_anyvector_ty], [llvm_anyvector_ty], [IntrNoMem]>;
-def int_aarch64_neon_fcvtzs :
+def int_aarch64_neon_vcvtzs :
   Intrinsic<[llvm_anyvector_ty], [llvm_anyvector_ty], [IntrNoMem]>;
-def int_aarch64_neon_fcvtzu :
+def int_aarch64_neon_vcvtzu :
   Intrinsic<[llvm_anyvector_ty], [llvm_anyvector_ty], [IntrNoMem]>;
 
 // Vector maxNum (Floating Point)
@@ -240,6 +240,30 @@ def int_aarch64_neon_vcvtf32_u32 :
 def int_aarch64_neon_vcvtf64_u64 :
   Intrinsic<[llvm_double_ty], [llvm_v1i64_ty], [IntrNoMem]>;
 
+// Scalar Floating-point Convert
+def int_aarch64_neon_fcvtxn :
+  Intrinsic<[llvm_float_ty], [llvm_double_ty], [IntrNoMem]>;
+def int_aarch64_neon_fcvtns : 
+  Intrinsic<[llvm_anyvector_ty], [llvm_anyfloat_ty], [IntrNoMem]>;
+def int_aarch64_neon_fcvtnu :
+  Intrinsic<[llvm_anyvector_ty], [llvm_anyfloat_ty], [IntrNoMem]>;
+def int_aarch64_neon_fcvtps :
+  Intrinsic<[llvm_anyvector_ty], [llvm_anyfloat_ty], [IntrNoMem]>;
+def int_aarch64_neon_fcvtpu :
+  Intrinsic<[llvm_anyvector_ty], [llvm_anyfloat_ty], [IntrNoMem]>;
+def int_aarch64_neon_fcvtms :
+  Intrinsic<[llvm_anyvector_ty], [llvm_anyfloat_ty], [IntrNoMem]>;
+def int_aarch64_neon_fcvtmu :
+  Intrinsic<[llvm_anyvector_ty], [llvm_anyfloat_ty], [IntrNoMem]>;
+def int_aarch64_neon_fcvtas :
+  Intrinsic<[llvm_anyvector_ty], [llvm_anyfloat_ty], [IntrNoMem]>;
+def int_aarch64_neon_fcvtau :
+  Intrinsic<[llvm_anyvector_ty], [llvm_anyfloat_ty], [IntrNoMem]>;
+def int_aarch64_neon_fcvtzs :
+  Intrinsic<[llvm_anyvector_ty], [llvm_anyfloat_ty], [IntrNoMem]>;
+def int_aarch64_neon_fcvtzu :
+  Intrinsic<[llvm_anyvector_ty], [llvm_anyfloat_ty], [IntrNoMem]>;
+
 // Scalar Floating-point Reciprocal Exponent
 def int_aarch64_neon_vrecpx : Neon_1Arg_Intrinsic;
 

llvm/lib/Target/AArch64/AArch64InstrNEON.td

@@ -4307,18 +4307,23 @@ multiclass NeonI_Scalar2SameMisc_accum_BHSD_size<bit u, bits<5> opcode,
 
 class Neon_Scalar2SameMisc_fcvtxn_D_size_patterns<SDPatternOperator opnode,
                                                   Instruction INSTD>
-  : Pat<(v1f32 (opnode (v1f64 FPR64:$Rn))),
+  : Pat<(f32 (opnode (f64 FPR64:$Rn))),
         (INSTD FPR64:$Rn)>;
 
 multiclass Neon_Scalar2SameMisc_fcvt_SD_size_patterns<SDPatternOperator opnode,
                                                       Instruction INSTS,
                                                       Instruction INSTD> {
-  def : Pat<(v1i32 (opnode (v1f32 FPR32:$Rn))),
+  def : Pat<(v1i32 (opnode (f32 FPR32:$Rn))),
             (INSTS FPR32:$Rn)>;
-  def : Pat<(v1i64 (opnode (v1f64 FPR64:$Rn))),
+  def : Pat<(v1i64 (opnode (f64 FPR64:$Rn))),
             (INSTD FPR64:$Rn)>;
 }
 
+class Neon_Scalar2SameMisc_vcvt_D_size_patterns<SDPatternOperator opnode,
+                                                Instruction INSTD>
+  : Pat<(v1i64 (opnode (v1f64 FPR64:$Rn))),
+            (INSTD FPR64:$Rn)>;
+
 multiclass Neon_Scalar2SameMisc_cvt_SD_size_patterns<SDPatternOperator Sopnode,
                                                      SDPatternOperator Dopnode,
                                                      Instruction INSTS,
@@ -4982,44 +4987,56 @@ def : Neon_Scalar2SameMisc_fcvtxn_D_size_patterns<int_aarch64_neon_fcvtxn,
                                                   FCVTXN>;
 
 defm FCVTNS : NeonI_Scalar2SameMisc_SD_size<0b0, 0b0, 0b11010, "fcvtns">;
-defm : Neon_Scalar2SameMisc_fcvt_SD_size_patterns<int_arm_neon_vcvtns,
+defm : Neon_Scalar2SameMisc_fcvt_SD_size_patterns<int_aarch64_neon_fcvtns,
                                                   FCVTNSss, FCVTNSdd>;
+def : Neon_Scalar2SameMisc_vcvt_D_size_patterns<int_arm_neon_vcvtns, FCVTNSdd>;
 
 defm FCVTNU : NeonI_Scalar2SameMisc_SD_size<0b1, 0b0, 0b11010, "fcvtnu">;
-defm : Neon_Scalar2SameMisc_fcvt_SD_size_patterns<int_arm_neon_vcvtnu,
+defm : Neon_Scalar2SameMisc_fcvt_SD_size_patterns<int_aarch64_neon_fcvtnu,
                                                   FCVTNUss, FCVTNUdd>;
+def : Neon_Scalar2SameMisc_vcvt_D_size_patterns<int_arm_neon_vcvtnu, FCVTNUdd>;
 
 defm FCVTMS : NeonI_Scalar2SameMisc_SD_size<0b0, 0b0, 0b11011, "fcvtms">;
-defm : Neon_Scalar2SameMisc_fcvt_SD_size_patterns<int_arm_neon_vcvtms,
+defm : Neon_Scalar2SameMisc_fcvt_SD_size_patterns<int_aarch64_neon_fcvtms,
                                                   FCVTMSss, FCVTMSdd>;
+def : Neon_Scalar2SameMisc_vcvt_D_size_patterns<int_arm_neon_vcvtms, FCVTMSdd>;
 
 defm FCVTMU : NeonI_Scalar2SameMisc_SD_size<0b1, 0b0, 0b11011, "fcvtmu">;
-defm : Neon_Scalar2SameMisc_fcvt_SD_size_patterns<int_arm_neon_vcvtmu,
+defm : Neon_Scalar2SameMisc_fcvt_SD_size_patterns<int_aarch64_neon_fcvtmu,
                                                   FCVTMUss, FCVTMUdd>;
+def : Neon_Scalar2SameMisc_vcvt_D_size_patterns<int_arm_neon_vcvtmu, FCVTMUdd>;
 
 defm FCVTAS : NeonI_Scalar2SameMisc_SD_size<0b0, 0b0, 0b11100, "fcvtas">;
-defm : Neon_Scalar2SameMisc_fcvt_SD_size_patterns<int_arm_neon_vcvtas,
+defm : Neon_Scalar2SameMisc_fcvt_SD_size_patterns<int_aarch64_neon_fcvtas,
                                                   FCVTASss, FCVTASdd>;
+def : Neon_Scalar2SameMisc_vcvt_D_size_patterns<int_arm_neon_vcvtas, FCVTASdd>;
 
 defm FCVTAU : NeonI_Scalar2SameMisc_SD_size<0b1, 0b0, 0b11100, "fcvtau">;
-defm : Neon_Scalar2SameMisc_fcvt_SD_size_patterns<int_arm_neon_vcvtau,
+defm : Neon_Scalar2SameMisc_fcvt_SD_size_patterns<int_aarch64_neon_fcvtau,
                                                   FCVTAUss, FCVTAUdd>;
+def : Neon_Scalar2SameMisc_vcvt_D_size_patterns<int_arm_neon_vcvtau, FCVTAUdd>;
 
 defm FCVTPS : NeonI_Scalar2SameMisc_SD_size<0b0, 0b1, 0b11010, "fcvtps">;
-defm : Neon_Scalar2SameMisc_fcvt_SD_size_patterns<int_arm_neon_vcvtps,
+defm : Neon_Scalar2SameMisc_fcvt_SD_size_patterns<int_aarch64_neon_fcvtps,
                                                   FCVTPSss, FCVTPSdd>;
+def : Neon_Scalar2SameMisc_vcvt_D_size_patterns<int_arm_neon_vcvtps, FCVTPSdd>;
 
 defm FCVTPU : NeonI_Scalar2SameMisc_SD_size<0b1, 0b1, 0b11010, "fcvtpu">;
-defm : Neon_Scalar2SameMisc_fcvt_SD_size_patterns<int_arm_neon_vcvtpu,
+defm : Neon_Scalar2SameMisc_fcvt_SD_size_patterns<int_aarch64_neon_fcvtpu,
                                                   FCVTPUss, FCVTPUdd>;
+def : Neon_Scalar2SameMisc_vcvt_D_size_patterns<int_arm_neon_vcvtpu, FCVTPUdd>;
 
 defm FCVTZS : NeonI_Scalar2SameMisc_SD_size<0b0, 0b1, 0b11011, "fcvtzs">;
 defm : Neon_Scalar2SameMisc_fcvt_SD_size_patterns<int_aarch64_neon_fcvtzs,
                                                   FCVTZSss, FCVTZSdd>;
+def : Neon_Scalar2SameMisc_vcvt_D_size_patterns<int_aarch64_neon_vcvtzs,
+                                                FCVTZSdd>;
 
 defm FCVTZU : NeonI_Scalar2SameMisc_SD_size<0b1, 0b1, 0b11011, "fcvtzu">;
 defm : Neon_Scalar2SameMisc_fcvt_SD_size_patterns<int_aarch64_neon_fcvtzu,
                                                   FCVTZUss, FCVTZUdd>;
+def : Neon_Scalar2SameMisc_vcvt_D_size_patterns<int_aarch64_neon_vcvtzu,
+                                                FCVTZUdd>;
 
 // Patterns For Convert Instructions Between v1f64 and v1i64
 class Neon_Scalar2SameMisc_cvtf_v1f64_pattern<SDPatternOperator opnode,
@@ -8297,12 +8314,12 @@ multiclass NeonI_2VMisc_D_Narrow<string asmop, string prefix, bit U,
     let Constraints = "$src = $Rd";
   }
 
-  def : Pat<(v2f32 (int_aarch64_neon_fcvtxn (v2f64 VPR128:$Rn))),
+  def : Pat<(v2f32 (int_aarch64_neon_vcvtxn (v2f64 VPR128:$Rn))),
             (!cast<Instruction>(prefix # "2d2s") VPR128:$Rn)>;
 
   def : Pat<(v4f32 (concat_vectors
               (v2f32 VPR64:$src),
-              (v2f32 (int_aarch64_neon_fcvtxn (v2f64 VPR128:$Rn))))),
+              (v2f32 (int_aarch64_neon_vcvtxn (v2f64 VPR128:$Rn))))),
             (!cast<Instruction>(prefix # "2d4s")
                (v4f32 (SUBREG_TO_REG (i32 0), VPR64:$src, sub_64)),
                VPR128:$Rn)>;

llvm/test/CodeGen/AArch64/neon-misc.ll

@@ -894,13 +894,13 @@ define <4 x float> @test_vcvt_high_f32_f64(<2 x float> %a, <2 x double> %b) #0 {
 
 define <2 x float> @test_vcvtx_f32_f64(<2 x double> %a) #0 {
 ; CHECK: fcvtxn v{{[0-9]+}}.2s, v{{[0-9]+}}.2d
-  %vcvtx_f32_f641.i = tail call <2 x float> @llvm.aarch64.neon.fcvtxn.v2f32.v2f64(<2 x double> %a) #4
+  %vcvtx_f32_f641.i = call <2 x float> @llvm.aarch64.neon.vcvtxn.v2f32.v2f64(<2 x double> %a) #4
   ret <2 x float> %vcvtx_f32_f641.i
 }
 
 define <4 x float> @test_vcvtx_high_f32_f64(<2 x float> %a, <2 x double> %b) #0 {
 ; CHECK: fcvtxn2 v{{[0-9]+}}.4s, v{{[0-9]+}}.2d
-  %vcvtx_f32_f641.i.i = tail call <2 x float> @llvm.aarch64.neon.fcvtxn.v2f32.v2f64(<2 x double> %b) #4
+  %vcvtx_f32_f641.i.i = tail call <2 x float> @llvm.aarch64.neon.vcvtxn.v2f32.v2f64(<2 x double> %b) #4
   %shuffle.i = shufflevector <2 x float> %a, <2 x float> %vcvtx_f32_f641.i.i, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
   ret <4 x float> %shuffle.i
 }
@@ -1462,7 +1462,7 @@ declare <4 x float> @llvm.aarch64.neon.frintn.v4f32(<4 x float>) #2
 
 declare <2 x float> @llvm.aarch64.neon.frintn.v2f32(<2 x float>) #2
 
-declare <2 x float> @llvm.aarch64.neon.fcvtxn.v2f32.v2f64(<2 x double>) #2
+declare <2 x float> @llvm.aarch64.neon.vcvtxn.v2f32.v2f64(<2 x double>) #2
 
 declare <2 x float> @llvm.aarch64.neon.fcvtn.v2f32.v2f64(<2 x double>) #2
 

llvm/test/CodeGen/AArch64/neon-scalar-fcvt.ll

@@ -6,250 +6,228 @@ define float @test_vcvtxn(double %a) {
 ; CHECK: test_vcvtxn
 ; CHECK: fcvtxn {{s[0-9]}}, {{d[0-9]}}
 entry:
-  %vcvtf.i = insertelement <1 x double> undef, double %a, i32 0
+  %vcvtf = call float @llvm.aarch64.neon.fcvtxn(double %a)
-  %vcvtf1.i = tail call <1 x float> @llvm.aarch64.neon.fcvtxn.v1f32.v1f64(<1 x double> %vcvtf.i)
+  ret float %vcvtf
-  %0 = extractelement <1 x float> %vcvtf1.i, i32 0
-  ret float %0
 }
 
-declare <1 x float> @llvm.aarch64.neon.fcvtxn.v1f32.v1f64(<1 x double>)
+declare float @llvm.aarch64.neon.fcvtxn(double)
 
 define i32 @test_vcvtass(float %a) {
 ; CHECK: test_vcvtass
 ; CHECK: fcvtas {{s[0-9]}}, {{s[0-9]}}
 entry:
-  %vcvtas.i = insertelement <1 x float> undef, float %a, i32 0
+  %vcvtas1.i = call <1 x i32> @llvm.aarch64.neon.fcvtas.v1i32.f32(float %a)
-  %vcvtas1.i = call <1 x i32> @llvm.arm.neon.vcvtas.v1i32.v1f32(<1 x float> %vcvtas.i)
   %0 = extractelement <1 x i32> %vcvtas1.i, i32 0
   ret i32 %0
 }
 
-declare <1 x i32> @llvm.arm.neon.vcvtas.v1i32.v1f32(<1 x float>)
+declare <1 x i32> @llvm.aarch64.neon.fcvtas.v1i32.f32(float)
 
 define i64 @test_test_vcvtasd(double %a) {
 ; CHECK: test_test_vcvtasd
 ; CHECK: fcvtas {{d[0-9]}}, {{d[0-9]}}
 entry:
-  %vcvtas.i = insertelement <1 x double> undef, double %a, i32 0
+  %vcvtas1.i = call <1 x i64> @llvm.aarch64.neon.fcvtas.v1i64.f64(double %a)
-  %vcvtas1.i = call <1 x i64> @llvm.arm.neon.vcvtas.v1i64.v1f64(<1 x double> %vcvtas.i)
   %0 = extractelement <1 x i64> %vcvtas1.i, i32 0
   ret i64 %0
 }
 
-declare <1 x i64> @llvm.arm.neon.vcvtas.v1i64.v1f64(<1 x double>)
+declare <1 x i64> @llvm.aarch64.neon.fcvtas.v1i64.f64(double)
 
 define i32 @test_vcvtaus(float %a) {
 ; CHECK: test_vcvtaus
 ; CHECK: fcvtau {{s[0-9]}}, {{s[0-9]}}
 entry:
-  %vcvtau.i = insertelement <1 x float> undef, float %a, i32 0
+  %vcvtau1.i = call <1 x i32> @llvm.aarch64.neon.fcvtau.v1i32.f32(float %a)
-  %vcvtau1.i = call <1 x i32> @llvm.arm.neon.vcvtau.v1i32.v1f32(<1 x float> %vcvtau.i)
   %0 = extractelement <1 x i32> %vcvtau1.i, i32 0
   ret i32 %0
 }
 
-declare <1 x i32> @llvm.arm.neon.vcvtau.v1i32.v1f32(<1 x float>)
+declare <1 x i32> @llvm.aarch64.neon.fcvtau.v1i32.f32(float)
 
 define i64 @test_vcvtaud(double %a) {
 ; CHECK: test_vcvtaud
 ; CHECK: fcvtau {{d[0-9]}}, {{d[0-9]}}
 entry:
-  %vcvtau.i = insertelement <1 x double> undef, double %a, i32 0
+  %vcvtau1.i = call <1 x i64> @llvm.aarch64.neon.fcvtau.v1i64.f64(double %a)
-  %vcvtau1.i = call <1 x i64> @llvm.arm.neon.vcvtau.v1i64.v1f64(<1 x double> %vcvtau.i)
   %0 = extractelement <1 x i64> %vcvtau1.i, i32 0
   ret i64 %0
 }
 
-declare <1 x i64> @llvm.arm.neon.vcvtau.v1i64.v1f64(<1 x double>) 
+declare <1 x i64> @llvm.aarch64.neon.fcvtau.v1i64.f64(double) 
 
 define i32 @test_vcvtmss(float %a) {
 ; CHECK: test_vcvtmss
 ; CHECK: fcvtms {{s[0-9]}}, {{s[0-9]}}
 entry:
-  %vcvtms.i = insertelement <1 x float> undef, float %a, i32 0
+  %vcvtms1.i = call <1 x i32> @llvm.aarch64.neon.fcvtms.v1i32.f32(float %a)
-  %vcvtms1.i = call <1 x i32> @llvm.arm.neon.vcvtms.v1i32.v1f32(<1 x float> %vcvtms.i)
   %0 = extractelement <1 x i32> %vcvtms1.i, i32 0
   ret i32 %0
 }
 
-declare <1 x i32> @llvm.arm.neon.vcvtms.v1i32.v1f32(<1 x float>)
+declare <1 x i32> @llvm.aarch64.neon.fcvtms.v1i32.f32(float)
 
 define i64 @test_vcvtmd_s64_f64(double %a) {
 ; CHECK: test_vcvtmd_s64_f64
 ; CHECK: fcvtms {{d[0-9]}}, {{d[0-9]}}
 entry:
-  %vcvtms.i = insertelement <1 x double> undef, double %a, i32 0
+  %vcvtms1.i = call <1 x i64> @llvm.aarch64.neon.fcvtms.v1i64.f64(double %a)
-  %vcvtms1.i = call <1 x i64> @llvm.arm.neon.vcvtms.v1i64.v1f64(<1 x double> %vcvtms.i)
   %0 = extractelement <1 x i64> %vcvtms1.i, i32 0
   ret i64 %0
 }
 
-declare <1 x i64> @llvm.arm.neon.vcvtms.v1i64.v1f64(<1 x double>)
+declare <1 x i64> @llvm.aarch64.neon.fcvtms.v1i64.f64(double)
 
 define i32 @test_vcvtmus(float %a) {
 ; CHECK: test_vcvtmus
 ; CHECK: fcvtmu {{s[0-9]}}, {{s[0-9]}}
 entry:
-  %vcvtmu.i = insertelement <1 x float> undef, float %a, i32 0
+  %vcvtmu1.i = call <1 x i32> @llvm.aarch64.neon.fcvtmu.v1i32.f32(float %a)
-  %vcvtmu1.i = call <1 x i32> @llvm.arm.neon.vcvtmu.v1i32.v1f32(<1 x float> %vcvtmu.i)
   %0 = extractelement <1 x i32> %vcvtmu1.i, i32 0
   ret i32 %0
 }
 
-declare <1 x i32> @llvm.arm.neon.vcvtmu.v1i32.v1f32(<1 x float>)
+declare <1 x i32> @llvm.aarch64.neon.fcvtmu.v1i32.f32(float)
 
 define i64 @test_vcvtmud(double %a) {
 ; CHECK: test_vcvtmud
 ; CHECK: fcvtmu {{d[0-9]}}, {{d[0-9]}}
 entry:
-  %vcvtmu.i = insertelement <1 x double> undef, double %a, i32 0
+  %vcvtmu1.i = call <1 x i64> @llvm.aarch64.neon.fcvtmu.v1i64.f64(double %a)
-  %vcvtmu1.i = call <1 x i64> @llvm.arm.neon.vcvtmu.v1i64.v1f64(<1 x double> %vcvtmu.i)
   %0 = extractelement <1 x i64> %vcvtmu1.i, i32 0
   ret i64 %0
 }
 
-declare <1 x i64> @llvm.arm.neon.vcvtmu.v1i64.v1f64(<1 x double>)
+declare <1 x i64> @llvm.aarch64.neon.fcvtmu.v1i64.f64(double)
 
 define i32 @test_vcvtnss(float %a) {
 ; CHECK: test_vcvtnss
 ; CHECK: fcvtns {{s[0-9]}}, {{s[0-9]}}
 entry:
-  %vcvtns.i = insertelement <1 x float> undef, float %a, i32 0
+  %vcvtns1.i = call <1 x i32> @llvm.aarch64.neon.fcvtns.v1i32.f32(float %a)
-  %vcvtns1.i = call <1 x i32> @llvm.arm.neon.vcvtns.v1i32.v1f32(<1 x float> %vcvtns.i)
   %0 = extractelement <1 x i32> %vcvtns1.i, i32 0
   ret i32 %0
 }
 
-declare <1 x i32> @llvm.arm.neon.vcvtns.v1i32.v1f32(<1 x float>)
+declare <1 x i32> @llvm.aarch64.neon.fcvtns.v1i32.f32(float)
 
 define i64 @test_vcvtnd_s64_f64(double %a) {
 ; CHECK: test_vcvtnd_s64_f64
 ; CHECK: fcvtns {{d[0-9]}}, {{d[0-9]}}
 entry:
-  %vcvtns.i = insertelement <1 x double> undef, double %a, i32 0
+  %vcvtns1.i = call <1 x i64> @llvm.aarch64.neon.fcvtns.v1i64.f64(double %a)
-  %vcvtns1.i = call <1 x i64> @llvm.arm.neon.vcvtns.v1i64.v1f64(<1 x double> %vcvtns.i)
   %0 = extractelement <1 x i64> %vcvtns1.i, i32 0
   ret i64 %0
 }
 
-declare <1 x i64> @llvm.arm.neon.vcvtns.v1i64.v1f64(<1 x double>)
+declare <1 x i64> @llvm.aarch64.neon.fcvtns.v1i64.f64(double)
 
 define i32 @test_vcvtnus(float %a) {
 ; CHECK: test_vcvtnus
 ; CHECK: fcvtnu {{s[0-9]}}, {{s[0-9]}}
 entry:
-  %vcvtnu.i = insertelement <1 x float> undef, float %a, i32 0
+  %vcvtnu1.i = call <1 x i32> @llvm.aarch64.neon.fcvtnu.v1i32.f32(float %a)
-  %vcvtnu1.i = call <1 x i32> @llvm.arm.neon.vcvtnu.v1i32.v1f32(<1 x float> %vcvtnu.i)
   %0 = extractelement <1 x i32> %vcvtnu1.i, i32 0
   ret i32 %0
 }
 
-declare <1 x i32> @llvm.arm.neon.vcvtnu.v1i32.v1f32(<1 x float>)
+declare <1 x i32> @llvm.aarch64.neon.fcvtnu.v1i32.f32(float)
 
 define i64 @test_vcvtnud(double %a) {
 ; CHECK: test_vcvtnud
 ; CHECK: fcvtnu {{d[0-9]}}, {{d[0-9]}}
 entry:
-  %vcvtnu.i = insertelement <1 x double> undef, double %a, i32 0
+  %vcvtnu1.i = call <1 x i64> @llvm.aarch64.neon.fcvtnu.v1i64.f64(double %a)
-  %vcvtnu1.i = call <1 x i64> @llvm.arm.neon.vcvtnu.v1i64.v1f64(<1 x double> %vcvtnu.i)
   %0 = extractelement <1 x i64> %vcvtnu1.i, i32 0
   ret i64 %0
 }
 
-declare <1 x i64> @llvm.arm.neon.vcvtnu.v1i64.v1f64(<1 x double>)
+declare <1 x i64> @llvm.aarch64.neon.fcvtnu.v1i64.f64(double)
 
 define i32 @test_vcvtpss(float %a) {
 ; CHECK: test_vcvtpss
 ; CHECK: fcvtps {{s[0-9]}}, {{s[0-9]}}
 entry:
-  %vcvtps.i = insertelement <1 x float> undef, float %a, i32 0
+  %vcvtps1.i = call <1 x i32> @llvm.aarch64.neon.fcvtps.v1i32.f32(float %a)
-  %vcvtps1.i = call <1 x i32> @llvm.arm.neon.vcvtps.v1i32.v1f32(<1 x float> %vcvtps.i)
   %0 = extractelement <1 x i32> %vcvtps1.i, i32 0
   ret i32 %0
 }
 
-declare <1 x i32> @llvm.arm.neon.vcvtps.v1i32.v1f32(<1 x float>)
+declare <1 x i32> @llvm.aarch64.neon.fcvtps.v1i32.f32(float)
 
 define i64 @test_vcvtpd_s64_f64(double %a) {
 ; CHECK: test_vcvtpd_s64_f64
 ; CHECK: fcvtps {{d[0-9]}}, {{d[0-9]}}
 entry:
-  %vcvtps.i = insertelement <1 x double> undef, double %a, i32 0
+  %vcvtps1.i = call <1 x i64> @llvm.aarch64.neon.fcvtps.v1i64.f64(double %a)
-  %vcvtps1.i = call <1 x i64> @llvm.arm.neon.vcvtps.v1i64.v1f64(<1 x double> %vcvtps.i)
   %0 = extractelement <1 x i64> %vcvtps1.i, i32 0
   ret i64 %0
 }
 
-declare <1 x i64> @llvm.arm.neon.vcvtps.v1i64.v1f64(<1 x double>)
+declare <1 x i64> @llvm.aarch64.neon.fcvtps.v1i64.f64(double)
 
 define i32 @test_vcvtpus(float %a) {
 ; CHECK: test_vcvtpus
 ; CHECK: fcvtpu {{s[0-9]}}, {{s[0-9]}}
 entry:
-  %vcvtpu.i = insertelement <1 x float> undef, float %a, i32 0
+  %vcvtpu1.i = call <1 x i32> @llvm.aarch64.neon.fcvtpu.v1i32.f32(float %a)
-  %vcvtpu1.i = call <1 x i32> @llvm.arm.neon.vcvtpu.v1i32.v1f32(<1 x float> %vcvtpu.i)
   %0 = extractelement <1 x i32> %vcvtpu1.i, i32 0
   ret i32 %0
 }
 
-declare <1 x i32> @llvm.arm.neon.vcvtpu.v1i32.v1f32(<1 x float>)
+declare <1 x i32> @llvm.aarch64.neon.fcvtpu.v1i32.f32(float)
 
 define i64 @test_vcvtpud(double %a) {
 ; CHECK: test_vcvtpud
 ; CHECK: fcvtpu {{d[0-9]}}, {{d[0-9]}}
 entry:
-  %vcvtpu.i = insertelement <1 x double> undef, double %a, i32 0
+  %vcvtpu1.i = call <1 x i64> @llvm.aarch64.neon.fcvtpu.v1i64.f64(double %a)
-  %vcvtpu1.i = tail call <1 x i64> @llvm.arm.neon.vcvtpu.v1i64.v1f64(<1 x double> %vcvtpu.i)
   %0 = extractelement <1 x i64> %vcvtpu1.i, i32 0
   ret i64 %0
 }
 
-declare <1 x i64> @llvm.arm.neon.vcvtpu.v1i64.v1f64(<1 x double>)
+declare <1 x i64> @llvm.aarch64.neon.fcvtpu.v1i64.f64(double)
 
 define i32 @test_vcvtss(float %a) {
 ; CHECK: test_vcvtss
 ; CHECK: fcvtzs {{s[0-9]}}, {{s[0-9]}}
 entry:
-  %vcvtzs.i = insertelement <1 x float> undef, float %a, i32 0
+  %vcvtzs1.i = call <1 x i32> @llvm.aarch64.neon.fcvtzs.v1i32.f32(float %a)
-  %vcvtzs1.i = tail call <1 x i32> @llvm.aarch64.neon.fcvtzs.v1i32.v1f32(<1 x float> %vcvtzs.i)
   %0 = extractelement <1 x i32> %vcvtzs1.i, i32 0
   ret i32 %0
 }
 
-declare <1 x i32> @llvm.aarch64.neon.fcvtzs.v1i32.v1f32(<1 x float>)
+declare <1 x i32> @llvm.aarch64.neon.fcvtzs.v1i32.f32(float)
 
 define i64 @test_vcvtd_s64_f64(double %a) {
 ; CHECK: test_vcvtd_s64_f64
 ; CHECK: fcvtzs {{d[0-9]}}, {{d[0-9]}}
 entry:
-  %vcvzs.i = insertelement <1 x double> undef, double %a, i32 0
+  %vcvzs1.i = call <1 x i64> @llvm.aarch64.neon.fcvtzs.v1i64.f64(double %a)
-  %vcvzs1.i = tail call <1 x i64> @llvm.aarch64.neon.fcvtzs.v1i64.v1f64(<1 x double> %vcvzs.i)
   %0 = extractelement <1 x i64> %vcvzs1.i, i32 0
   ret i64 %0
 }
 
-declare <1 x i64> @llvm.aarch64.neon.fcvtzs.v1i64.v1f64(<1 x double>)
+declare <1 x i64> @llvm.aarch64.neon.fcvtzs.v1i64.f64(double)
 
 define i32 @test_vcvtus(float %a) {
 ; CHECK: test_vcvtus
 ; CHECK: fcvtzu {{s[0-9]}}, {{s[0-9]}}
 entry:
-  %vcvtzu.i = insertelement <1 x float> undef, float %a, i32 0
+  %vcvtzu1.i = call <1 x i32> @llvm.aarch64.neon.fcvtzu.v1i32.f32(float %a)
-  %vcvtzu1.i = tail call <1 x i32> @llvm.aarch64.neon.fcvtzu.v1i32.v1f32(<1 x float> %vcvtzu.i)
   %0 = extractelement <1 x i32> %vcvtzu1.i, i32 0
   ret i32 %0
 }
 
-declare <1 x i32> @llvm.aarch64.neon.fcvtzu.v1i32.v1f32(<1 x float>)
+declare <1 x i32> @llvm.aarch64.neon.fcvtzu.v1i32.f32(float)
 
 define i64 @test_vcvtud(double %a) {
 ; CHECK: test_vcvtud
 ; CHECK: fcvtzu {{d[0-9]}}, {{d[0-9]}}
 entry:
-  %vcvtzu.i = insertelement <1 x double> undef, double %a, i32 0
+  %vcvtzu1.i = call <1 x i64> @llvm.aarch64.neon.fcvtzu.v1i64.f64(double %a)
-  %vcvtzu1.i = tail call <1 x i64> @llvm.aarch64.neon.fcvtzu.v1i64.v1f64(<1 x double> %vcvtzu.i)
   %0 = extractelement <1 x i64> %vcvtzu1.i, i32 0
   ret i64 %0
 }
 
-declare <1 x i64> @llvm.aarch64.neon.fcvtzu.v1i64.v1f64(<1 x double>)
+declare <1 x i64> @llvm.aarch64.neon.fcvtzu.v1i64.f64(double)