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>, <4 x double>* %0, align 32 14 %1 = bitcast float* %f to <8 x float>* 15 %tmp1.i17 = load <8 x float>, <8 x float>* %1, align 32 16 %tmp1.i16 = load <4 x i64>, <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 or vinsertps from memory 29 30; CHECK: mov00 31define <8 x float> @mov00(<8 x float> %v, float * %ptr) nounwind { 32 %val = load float, float* %ptr 33; CHECK: vmovss (% 34 %i0 = insertelement <8 x float> zeroinitializer, float %val, i32 0 35 ret <8 x float> %i0 36; CHECK: ret 37} 38 39; CHECK: mov01 40define <4 x double> @mov01(<4 x double> %v, double * %ptr) nounwind { 41 %val = load double, double* %ptr 42; CHECK: vmovsd (% 43 %i0 = insertelement <4 x double> zeroinitializer, double %val, i32 0 44 ret <4 x double> %i0 45; CHECK: ret 46} 47 48; CHECK: vmovaps %ymm 49define void @storev16i16(<16 x i16> %a) nounwind { 50 store <16 x i16> %a, <16 x i16>* undef, align 32 51 unreachable 52} 53 54; CHECK: storev16i16_01 55; CHECK: vextractf128 56; CHECK: vmovups %xmm 57define void @storev16i16_01(<16 x i16> %a) nounwind { 58 store <16 x i16> %a, <16 x i16>* undef, align 4 59 unreachable 60} 61 62; CHECK: storev32i8 63; CHECK: vmovaps %ymm 64define void @storev32i8(<32 x i8> %a) nounwind { 65 store <32 x i8> %a, <32 x i8>* undef, align 32 66 unreachable 67} 68 69; CHECK: storev32i8_01 70; CHECK: vextractf128 71; CHECK: vmovups %xmm 72define void @storev32i8_01(<32 x i8> %a) nounwind { 73 store <32 x i8> %a, <32 x i8>* undef, align 4 74 unreachable 75} 76 77; It is faster to make two saves, if the data is already in XMM registers. For 78; example, after making an integer operation. 79; CHECK: _double_save 80; CHECK-NOT: vinsertf128 $1 81; CHECK-NOT: vinsertf128 $0 82; CHECK: vmovaps %xmm 83; CHECK: vmovaps %xmm 84define void @double_save(<4 x i32> %A, <4 x i32> %B, <8 x i32>* %P) nounwind ssp { 85entry: 86 %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> 87 store <8 x i32> %Z, <8 x i32>* %P, align 16 88 ret void 89} 90 91declare void @llvm.x86.avx.maskstore.ps.256(i8*, <8 x i32>, <8 x float>) nounwind 92 93; CHECK_O0: _f_f 94; CHECK-O0: vmovss LCPI 95; CHECK-O0: vxorps %xmm 96; CHECK-O0: vmovss %xmm 97define void @f_f() nounwind { 98allocas: 99 br i1 undef, label %cif_mask_all, label %cif_mask_mixed 100 101cif_mask_all: ; preds = %allocas 102 unreachable 103 104cif_mask_mixed: ; preds = %allocas 105 br i1 undef, label %cif_mixed_test_all, label %cif_mixed_test_any_check 106 107cif_mixed_test_all: ; preds = %cif_mask_mixed 108 call void @llvm.x86.avx.maskstore.ps.256(i8* undef, <8 x i32> <i32 -1, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>, <8 x float> undef) nounwind 109 unreachable 110 111cif_mixed_test_any_check: ; preds = %cif_mask_mixed 112 unreachable 113} 114 115; CHECK: add8i32 116; CHECK: vmovups 117; CHECK: vmovups 118; CHECK-NOT: vinsertf128 119; CHECK-NOT: vextractf128 120; CHECK: vmovups 121; CHECK: vmovups 122define void @add8i32(<8 x i32>* %ret, <8 x i32>* %bp) nounwind { 123 %b = load <8 x i32>, <8 x i32>* %bp, align 1 124 %x = add <8 x i32> zeroinitializer, %b 125 store <8 x i32> %x, <8 x i32>* %ret, align 1 126 ret void 127} 128 129; CHECK: add4i64a64 130; CHECK: vmovaps ({{.*}}), %ymm{{.*}} 131; CHECK: vmovaps %ymm{{.*}}, ({{.*}}) 132define void @add4i64a64(<4 x i64>* %ret, <4 x i64>* %bp) nounwind { 133 %b = load <4 x i64>, <4 x i64>* %bp, align 64 134 %x = add <4 x i64> zeroinitializer, %b 135 store <4 x i64> %x, <4 x i64>* %ret, align 64 136 ret void 137} 138 139; CHECK: add4i64a16 140; CHECK: vmovaps {{.*}}({{.*}}), %xmm{{.*}} 141; CHECK: vmovaps {{.*}}({{.*}}), %xmm{{.*}} 142; CHECK: vmovaps %xmm{{.*}}, {{.*}}({{.*}}) 143; CHECK: vmovaps %xmm{{.*}}, {{.*}}({{.*}}) 144define void @add4i64a16(<4 x i64>* %ret, <4 x i64>* %bp) nounwind { 145 %b = load <4 x i64>, <4 x i64>* %bp, align 16 146 %x = add <4 x i64> zeroinitializer, %b 147 store <4 x i64> %x, <4 x i64>* %ret, align 16 148 ret void 149} 150