• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1; RUN: llc < %s -mtriple=aarch64-unknown-unknown -mcpu=cyclone -mattr=+slow-misaligned-128store | FileCheck %s --check-prefix=SPLITTING --check-prefix=CHECK
2; RUN: llc < %s -mtriple=aarch64-eabi -mattr=-slow-misaligned-128store | FileCheck %s --check-prefix=MISALIGNED --check-prefix=CHECK
3
4@g0 = external global <3 x float>, align 16
5@g1 = external global <3 x float>, align 4
6
7; CHECK: ldr q[[R0:[0-9]+]], {{\[}}[[R1:x[0-9]+]], :lo12:g0
8; CHECK: str d[[R0]]
9
10define void @blam() {
11  %tmp4 = getelementptr inbounds <3 x float>, <3 x float>* @g1, i64 0, i64 0
12  %tmp5 = load <3 x float>, <3 x float>* @g0, align 16
13  %tmp6 = extractelement <3 x float> %tmp5, i64 0
14  store float %tmp6, float* %tmp4
15  %tmp7 = getelementptr inbounds float, float* %tmp4, i64 1
16  %tmp8 = load <3 x float>, <3 x float>* @g0, align 16
17  %tmp9 = extractelement <3 x float> %tmp8, i64 1
18  store float %tmp9, float* %tmp7
19  ret void;
20}
21
22
23; PR21711 - Merge vector stores into wider vector stores.
24
25; On Cyclone, the stores should not get merged into a 16-byte store because
26; unaligned 16-byte stores are slow. This test would infinite loop when
27; the fastness of unaligned accesses was not specified correctly.
28
29define void @merge_vec_extract_stores(<4 x float> %v1, <2 x float>* %ptr) {
30  %idx0 = getelementptr inbounds <2 x float>, <2 x float>* %ptr, i64 3
31  %idx1 = getelementptr inbounds <2 x float>, <2 x float>* %ptr, i64 4
32
33  %shuffle0 = shufflevector <4 x float> %v1, <4 x float> undef, <2 x i32> <i32 0, i32 1>
34  %shuffle1 = shufflevector <4 x float> %v1, <4 x float> undef, <2 x i32> <i32 2, i32 3>
35
36  store <2 x float> %shuffle0, <2 x float>* %idx0, align 8
37  store <2 x float> %shuffle1, <2 x float>* %idx1, align 8
38  ret void
39
40; MISALIGNED-LABEL:    merge_vec_extract_stores
41; MISALIGNED:          stur   q0, [x0, #24]
42; MISALIGNED-NEXT:     ret
43
44; FIXME: Ideally we would like to use a generic target for this test, but this relies
45; on suppressing store pairs.
46
47; SPLITTING-LABEL:    merge_vec_extract_stores
48; SPLITTING:          ext   v1.16b, v0.16b, v0.16b, #8
49; SPLITTING-NEXT:     str   d0, [x0, #24]
50; SPLITTING-NEXT:     str   d1, [x0, #32]
51; SPLITTING-NEXT:     ret
52}
53