• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1; RUN: llc -march=x86-64 -mtriple=x86_64-unknown-unknown < %s | FileCheck -check-prefix=DEFAULTCPU -check-prefix=ALL %s
2; RUN: llc -march=x86-64 -mcpu=x86-64 -mtriple=x86_64-unknown-unknown < %s | FileCheck -check-prefix=X8664CPU -check-prefix=ALL %s
3
4
5; ALL-LABEL: {{^}}merge_8_float_zero_stores:
6
7; DEFAULTCPU-DAG: movq $0, ([[PTR:%[a-z]+]])
8; DEFAULTCPU-DAG: movq $0, 8([[PTR]])
9; DEFAULTCPU-DAG: movq $0, 16([[PTR]])
10; DEFAULTCPU-DAG: movq $0, 24([[PTR]])
11
12; X8664CPU: xorps [[ZEROREG:%xmm[0-9]+]], [[ZEROREG]]
13; X8664CPU-DAG: movups [[ZEROREG]], ([[PTR:%[a-z]+]])
14; X8664CPU-DAG: movups [[ZEROREG]], 16([[PTR:%[a-z]+]])
15
16; ALL: retq
17define void @merge_8_float_zero_stores(float* %ptr) {
18  %idx0 = getelementptr float, float* %ptr, i64 0
19  %idx1 = getelementptr float, float* %ptr, i64 1
20  %idx2 = getelementptr float, float* %ptr, i64 2
21  %idx3 = getelementptr float, float* %ptr, i64 3
22  %idx4 = getelementptr float, float* %ptr, i64 4
23  %idx5 = getelementptr float, float* %ptr, i64 5
24  %idx6 = getelementptr float, float* %ptr, i64 6
25  %idx7 = getelementptr float, float* %ptr, i64 7
26  store float 0.0, float* %idx0, align 4
27  store float 0.0, float* %idx1, align 4
28  store float 0.0, float* %idx2, align 4
29  store float 0.0, float* %idx3, align 4
30  store float 0.0, float* %idx4, align 4
31  store float 0.0, float* %idx5, align 4
32  store float 0.0, float* %idx6, align 4
33  store float 0.0, float* %idx7, align 4
34  ret void
35}
36