1target triple = "x86_64-unknown-unknown" 2 3; RUN: llc < %s -march=x86-64 -mattr=+avx | FileCheck %s 4 5; When extracting multiple consecutive elements from a larger 6; vector into a smaller one, do it efficiently. We should use 7; an EXTRACT_SUBVECTOR node internally rather than a bunch of 8; single element extractions. 9 10; Extracting the low elements only requires using the right kind of store. 11define void @low_v8f32_to_v4f32(<8 x float> %v, <4 x float>* %ptr) { 12 %ext0 = extractelement <8 x float> %v, i32 0 13 %ext1 = extractelement <8 x float> %v, i32 1 14 %ext2 = extractelement <8 x float> %v, i32 2 15 %ext3 = extractelement <8 x float> %v, i32 3 16 %ins0 = insertelement <4 x float> undef, float %ext0, i32 0 17 %ins1 = insertelement <4 x float> %ins0, float %ext1, i32 1 18 %ins2 = insertelement <4 x float> %ins1, float %ext2, i32 2 19 %ins3 = insertelement <4 x float> %ins2, float %ext3, i32 3 20 store <4 x float> %ins3, <4 x float>* %ptr, align 16 21 ret void 22 23; CHECK-LABEL: low_v8f32_to_v4f32 24; CHECK: vmovaps 25; CHECK-NEXT: vzeroupper 26; CHECK-NEXT: retq 27} 28 29; Extracting the high elements requires just one AVX instruction. 30define void @high_v8f32_to_v4f32(<8 x float> %v, <4 x float>* %ptr) { 31 %ext0 = extractelement <8 x float> %v, i32 4 32 %ext1 = extractelement <8 x float> %v, i32 5 33 %ext2 = extractelement <8 x float> %v, i32 6 34 %ext3 = extractelement <8 x float> %v, i32 7 35 %ins0 = insertelement <4 x float> undef, float %ext0, i32 0 36 %ins1 = insertelement <4 x float> %ins0, float %ext1, i32 1 37 %ins2 = insertelement <4 x float> %ins1, float %ext2, i32 2 38 %ins3 = insertelement <4 x float> %ins2, float %ext3, i32 3 39 store <4 x float> %ins3, <4 x float>* %ptr, align 16 40 ret void 41 42; CHECK-LABEL: high_v8f32_to_v4f32 43; CHECK: vextractf128 44; CHECK-NEXT: vzeroupper 45; CHECK-NEXT: retq 46} 47 48; Make sure element type doesn't alter the codegen. Note that 49; if we were actually using the vector in this function and 50; have AVX2, we should generate vextracti128 (the int version). 51define void @high_v8i32_to_v4i32(<8 x i32> %v, <4 x i32>* %ptr) { 52 %ext0 = extractelement <8 x i32> %v, i32 4 53 %ext1 = extractelement <8 x i32> %v, i32 5 54 %ext2 = extractelement <8 x i32> %v, i32 6 55 %ext3 = extractelement <8 x i32> %v, i32 7 56 %ins0 = insertelement <4 x i32> undef, i32 %ext0, i32 0 57 %ins1 = insertelement <4 x i32> %ins0, i32 %ext1, i32 1 58 %ins2 = insertelement <4 x i32> %ins1, i32 %ext2, i32 2 59 %ins3 = insertelement <4 x i32> %ins2, i32 %ext3, i32 3 60 store <4 x i32> %ins3, <4 x i32>* %ptr, align 16 61 ret void 62 63; CHECK-LABEL: high_v8i32_to_v4i32 64; CHECK: vextractf128 65; CHECK-NEXT: vzeroupper 66; CHECK-NEXT: retq 67} 68 69; Make sure that element size doesn't alter the codegen. 70define void @high_v4f64_to_v2f64(<4 x double> %v, <2 x double>* %ptr) { 71 %ext0 = extractelement <4 x double> %v, i32 2 72 %ext1 = extractelement <4 x double> %v, i32 3 73 %ins0 = insertelement <2 x double> undef, double %ext0, i32 0 74 %ins1 = insertelement <2 x double> %ins0, double %ext1, i32 1 75 store <2 x double> %ins1, <2 x double>* %ptr, align 16 76 ret void 77 78; CHECK-LABEL: high_v4f64_to_v2f64 79; CHECK: vextractf128 80; CHECK-NEXT: vzeroupper 81; CHECK-NEXT: retq 82} 83