173
|
1 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
|
|
2
|
|
3 @L = external local_unnamed_addr addrspace(3) global [9 x double], align 16
|
|
4 @Ldisp = external local_unnamed_addr addrspace(3) global [96 x double], align 16
|
|
5
|
|
6 ; Stores are reordered during loads merge. This case used to assert while
|
|
7 ; scanning for a paired instruction because it used to expect paired one
|
|
8 ; to follow a base one.
|
|
9
|
|
10 ; GCN-LABEL: {{^}}out_of_order_merge:
|
|
11 ; GCN-COUNT2: ds_read2_b64
|
|
12 ; GCN-COUNT3: ds_write_b64
|
|
13 define amdgpu_kernel void @out_of_order_merge() {
|
|
14 entry:
|
252
|
15 %gep2 = getelementptr inbounds [96 x double], ptr addrspace(3) @Ldisp, i32 0, i32 1
|
|
16 %tmp12 = load <2 x double>, ptr addrspace(3) getelementptr inbounds ([9 x double], ptr addrspace(3) @L, i32 0, i32 1), align 8
|
173
|
17 %tmp14 = extractelement <2 x double> %tmp12, i32 0
|
|
18 %tmp15 = extractelement <2 x double> %tmp12, i32 1
|
|
19 %add50.i = fadd double %tmp14, %tmp15
|
252
|
20 store double %add50.i, ptr addrspace(3) @Ldisp, align 8
|
|
21 %tmp16 = load double, ptr addrspace(3) getelementptr inbounds ([9 x double], ptr addrspace(3) @L, i32 1, i32 0), align 8
|
|
22 store double %tmp16, ptr addrspace(3) %gep2, align 8
|
|
23 %tmp17 = load <2 x double>, ptr addrspace(3) getelementptr inbounds ([9 x double], ptr addrspace(3) @L, i32 2, i32 1), align 8
|
173
|
24 %tmp19 = extractelement <2 x double> %tmp17, i32 1
|
252
|
25 store double %tmp19, ptr addrspace(3) undef, align 8
|
173
|
26 ret void
|
|
27 }
|