1; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=corei7-avx -mattr=+avx | FileCheck %s 2; RUN: llc -O0 < %s -mtriple=x86_64-apple-darwin -mcpu=corei7-avx -mattr=+avx | FileCheck %s -check-prefix=CHECK_O0 3 4; CHECK: vmovaps 5; CHECK: vmovaps 6; CHECK: vmovaps 7; CHECK: vmovaps 8; CHECK: vmovaps 9; CHECK: vmovaps 10define void @test_256_load(double* nocapture %d, float* nocapture %f, <4 x i64>* nocapture %i) nounwind uwtable ssp { 11entry: 12 %0 = bitcast double* %d to <4 x double>* 13 %tmp1.i = load <4 x double>* %0, align 32 14 %1 = bitcast float* %f to <8 x float>* 15 %tmp1.i17 = load <8 x float>* %1, align 32 16 %tmp1.i16 = load <4 x i64>* %i, align 32 17 tail call void @dummy(<4 x double> %tmp1.i, <8 x float> %tmp1.i17, <4 x i64> %tmp1.i16) nounwind 18 store <4 x double> %tmp1.i, <4 x double>* %0, align 32 19 store <8 x float> %tmp1.i17, <8 x float>* %1, align 32 20 store <4 x i64> %tmp1.i16, <4 x i64>* %i, align 32 21 ret void 22} 23 24declare void @dummy(<4 x double>, <8 x float>, <4 x i64>) 25 26;; 27;; The two tests below check that we must fold load + scalar_to_vector 28;; + ins_subvec+ zext into only a single vmovss or vmovsd 29 30; CHECK: vmovss (% 31define <8 x float> @mov00(<8 x float> %v, float * %ptr) nounwind { 32 %val = load float* %ptr 33 %i0 = insertelement <8 x float> zeroinitializer, float %val, i32 0 34 ret <8 x float> %i0 35} 36 37; CHECK: vmovsd (% 38define <4 x double> @mov01(<4 x double> %v, double * %ptr) nounwind { 39 %val = load double* %ptr 40 %i0 = insertelement <4 x double> zeroinitializer, double %val, i32 0 41 ret <4 x double> %i0 42} 43 44; CHECK: vmovaps %ymm 45define void @storev16i16(<16 x i16> %a) nounwind { 46 store <16 x i16> %a, <16 x i16>* undef, align 32 47 unreachable 48} 49 50; CHECK: vmovups %ymm 51define void @storev16i16_01(<16 x i16> %a) nounwind { 52 store <16 x i16> %a, <16 x i16>* undef, align 4 53 unreachable 54} 55 56; CHECK: vmovaps %ymm 57define void @storev32i8(<32 x i8> %a) nounwind { 58 store <32 x i8> %a, <32 x i8>* undef, align 32 59 unreachable 60} 61 62; CHECK: vmovups %ymm 63define void @storev32i8_01(<32 x i8> %a) nounwind { 64 store <32 x i8> %a, <32 x i8>* undef, align 4 65 unreachable 66} 67 68; It is faster to make two saves, if the data is already in XMM registers. For 69; example, after making an integer operation. 70; CHECK: _double_save 71; CHECK-NOT: vinsertf128 $1 72; CHECK-NOT: vinsertf128 $0 73; CHECK: vmovaps %xmm 74; CHECK: vmovaps %xmm 75define void @double_save(<4 x i32> %A, <4 x i32> %B, <8 x i32>* %P) nounwind ssp { 76entry: 77 %Z = shufflevector <4 x i32>%A, <4 x i32>%B, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7> 78 store <8 x i32> %Z, <8 x i32>* %P, align 16 79 ret void 80} 81 82declare void @llvm.x86.avx.maskstore.ps.256(i8*, <8 x float>, <8 x float>) nounwind 83 84; CHECK_O0: _f_f 85; CHECK-O0: vmovss LCPI 86; CHECK-O0: vxorps %xmm 87; CHECK-O0: vmovss %xmm 88define void @f_f() nounwind { 89allocas: 90 br i1 undef, label %cif_mask_all, label %cif_mask_mixed 91 92cif_mask_all: ; preds = %allocas 93 unreachable 94 95cif_mask_mixed: ; preds = %allocas 96 br i1 undef, label %cif_mixed_test_all, label %cif_mixed_test_any_check 97 98cif_mixed_test_all: ; preds = %cif_mask_mixed 99 call void @llvm.x86.avx.maskstore.ps.256(i8* undef, <8 x float> <float 0xFFFFFFFFE0000000, float 0.000000e+00, float 0.000000e+00, float 0.000000e+00, float 0.000000e+00, float 0.000000e+00, float 0.000000e+00, float 0.000000e+00>, <8 x float> undef) nounwind 100 unreachable 101 102cif_mixed_test_any_check: ; preds = %cif_mask_mixed 103 unreachable 104} 105 106