Home | History | Annotate | Download | only in AArch64
      1 ; RUN: llc -mtriple arm64-apple-ios -O3 -o - < %s | FileCheck %s
      2 ; <rdar://problem/14477220>
      3 
      4 %class.Complex = type { float, float }
      5 %class.Complex_int = type { i32, i32 }
      6 %class.Complex_long = type { i64, i64 }
      7 
      8 ; CHECK-LABEL: @test
      9 ; CHECK: add [[BASE:x[0-9]+]], x0, x1, lsl #3
     10 ; CHECK: ldp [[CPLX1_I:s[0-9]+]], [[CPLX1_R:s[0-9]+]], {{\[}}[[BASE]]]
     11 ; CHECK: ldp [[CPLX2_I:s[0-9]+]], [[CPLX2_R:s[0-9]+]], {{\[}}[[BASE]], #64]
     12 ; CHECK: fadd {{s[0-9]+}}, [[CPLX2_I]], [[CPLX1_I]]
     13 ; CHECK: fadd {{s[0-9]+}}, [[CPLX2_R]], [[CPLX1_R]]
     14 ; CHECK: ret
     15 define void @test(%class.Complex* nocapture %out, i64 %out_start) {
     16 entry:
     17   %arrayidx = getelementptr inbounds %class.Complex, %class.Complex* %out, i64 %out_start
     18   %0 = bitcast %class.Complex* %arrayidx to i64*
     19   %1 = load i64, i64* %0, align 4
     20   %t0.sroa.0.0.extract.trunc = trunc i64 %1 to i32
     21   %2 = bitcast i32 %t0.sroa.0.0.extract.trunc to float
     22   %t0.sroa.2.0.extract.shift = lshr i64 %1, 32
     23   %t0.sroa.2.0.extract.trunc = trunc i64 %t0.sroa.2.0.extract.shift to i32
     24   %3 = bitcast i32 %t0.sroa.2.0.extract.trunc to float
     25   %add = add i64 %out_start, 8
     26   %arrayidx2 = getelementptr inbounds %class.Complex, %class.Complex* %out, i64 %add
     27   %i.i = getelementptr inbounds %class.Complex, %class.Complex* %arrayidx2, i64 0, i32 0
     28   %4 = load float, float* %i.i, align 4
     29   %add.i = fadd float %4, %2
     30   %retval.sroa.0.0.vec.insert.i = insertelement <2 x float> undef, float %add.i, i32 0
     31   %r.i = getelementptr inbounds %class.Complex, %class.Complex* %arrayidx2, i64 0, i32 1
     32   %5 = load float, float* %r.i, align 4
     33   %add5.i = fadd float %5, %3
     34   %retval.sroa.0.4.vec.insert.i = insertelement <2 x float> %retval.sroa.0.0.vec.insert.i, float %add5.i, i32 1
     35   %ref.tmp.sroa.0.0.cast = bitcast %class.Complex* %arrayidx to <2 x float>*
     36   store <2 x float> %retval.sroa.0.4.vec.insert.i, <2 x float>* %ref.tmp.sroa.0.0.cast, align 4
     37   ret void
     38 }
     39 
     40 ; CHECK-LABEL: @test_int
     41 ; CHECK: add [[BASE:x[0-9]+]], x0, x1, lsl #3
     42 ; CHECK: ldp [[CPLX1_I:w[0-9]+]], [[CPLX1_R:w[0-9]+]], {{\[}}[[BASE]]]
     43 ; CHECK: ldp [[CPLX2_I:w[0-9]+]], [[CPLX2_R:w[0-9]+]], {{\[}}[[BASE]], #64]
     44 ; CHECK: add {{w[0-9]+}}, [[CPLX2_I]], [[CPLX1_I]]
     45 ; CHECK: add {{w[0-9]+}}, [[CPLX2_R]], [[CPLX1_R]]
     46 ; CHECK: ret
     47 define void @test_int(%class.Complex_int* nocapture %out, i64 %out_start) {
     48 entry:
     49   %arrayidx = getelementptr inbounds %class.Complex_int, %class.Complex_int* %out, i64 %out_start
     50   %0 = bitcast %class.Complex_int* %arrayidx to i64*
     51   %1 = load i64, i64* %0, align 4
     52   %t0.sroa.0.0.extract.trunc = trunc i64 %1 to i32
     53   %2 = bitcast i32 %t0.sroa.0.0.extract.trunc to i32
     54   %t0.sroa.2.0.extract.shift = lshr i64 %1, 32
     55   %t0.sroa.2.0.extract.trunc = trunc i64 %t0.sroa.2.0.extract.shift to i32
     56   %3 = bitcast i32 %t0.sroa.2.0.extract.trunc to i32
     57   %add = add i64 %out_start, 8
     58   %arrayidx2 = getelementptr inbounds %class.Complex_int, %class.Complex_int* %out, i64 %add
     59   %i.i = getelementptr inbounds %class.Complex_int, %class.Complex_int* %arrayidx2, i64 0, i32 0
     60   %4 = load i32, i32* %i.i, align 4
     61   %add.i = add i32 %4, %2
     62   %retval.sroa.0.0.vec.insert.i = insertelement <2 x i32> undef, i32 %add.i, i32 0
     63   %r.i = getelementptr inbounds %class.Complex_int, %class.Complex_int* %arrayidx2, i64 0, i32 1
     64   %5 = load i32, i32* %r.i, align 4
     65   %add5.i = add i32 %5, %3
     66   %retval.sroa.0.4.vec.insert.i = insertelement <2 x i32> %retval.sroa.0.0.vec.insert.i, i32 %add5.i, i32 1
     67   %ref.tmp.sroa.0.0.cast = bitcast %class.Complex_int* %arrayidx to <2 x i32>*
     68   store <2 x i32> %retval.sroa.0.4.vec.insert.i, <2 x i32>* %ref.tmp.sroa.0.0.cast, align 4
     69   ret void
     70 }
     71 
     72 ; CHECK-LABEL: @test_long
     73 ; CHECK: add [[BASE:x[0-9]+]], x0, x1, lsl #4
     74 ; CHECK: ldp [[CPLX1_I:x[0-9]+]], [[CPLX1_R:x[0-9]+]], {{\[}}[[BASE]]]
     75 ; CHECK: ldp [[CPLX2_I:x[0-9]+]], [[CPLX2_R:x[0-9]+]], {{\[}}[[BASE]], #128]
     76 ; CHECK: add {{x[0-9]+}}, [[CPLX2_I]], [[CPLX1_I]]
     77 ; CHECK: add {{x[0-9]+}}, [[CPLX2_R]], [[CPLX1_R]]
     78 ; CHECK: ret
     79 define void @test_long(%class.Complex_long* nocapture %out, i64 %out_start) {
     80 entry:
     81   %arrayidx = getelementptr inbounds %class.Complex_long, %class.Complex_long* %out, i64 %out_start
     82   %0 = bitcast %class.Complex_long* %arrayidx to i128*
     83   %1 = load i128, i128* %0, align 4
     84   %t0.sroa.0.0.extract.trunc = trunc i128 %1 to i64
     85   %2 = bitcast i64 %t0.sroa.0.0.extract.trunc to i64
     86   %t0.sroa.2.0.extract.shift = lshr i128 %1, 64
     87   %t0.sroa.2.0.extract.trunc = trunc i128 %t0.sroa.2.0.extract.shift to i64
     88   %3 = bitcast i64 %t0.sroa.2.0.extract.trunc to i64
     89   %add = add i64 %out_start, 8
     90   %arrayidx2 = getelementptr inbounds %class.Complex_long, %class.Complex_long* %out, i64 %add
     91   %i.i = getelementptr inbounds %class.Complex_long, %class.Complex_long* %arrayidx2, i32 0, i32 0
     92   %4 = load i64, i64* %i.i, align 4
     93   %add.i = add i64 %4, %2
     94   %retval.sroa.0.0.vec.insert.i = insertelement <2 x i64> undef, i64 %add.i, i32 0
     95   %r.i = getelementptr inbounds %class.Complex_long, %class.Complex_long* %arrayidx2, i32 0, i32 1
     96   %5 = load i64, i64* %r.i, align 4
     97   %add5.i = add i64 %5, %3
     98   %retval.sroa.0.4.vec.insert.i = insertelement <2 x i64> %retval.sroa.0.0.vec.insert.i, i64 %add5.i, i32 1
     99   %ref.tmp.sroa.0.0.cast = bitcast %class.Complex_long* %arrayidx to <2 x i64>*
    100   store <2 x i64> %retval.sroa.0.4.vec.insert.i, <2 x i64>* %ref.tmp.sroa.0.0.cast, align 4
    101   ret void
    102 }
    103