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