3 ; RUN: llc -march=r600 -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=SI %s
5 declare void @llvm.AMDGPU.barrier.local() noduplicate nounwind
7 ; SI-LABEL: @indirect_access_f64_alloca:
8 ; SI: BUFFER_STORE_DWORD
9 define void @f64_alloca(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in, i32 %b) nounwind {
10 %val = load double addrspace(1)* %in, align 8
11 %array = alloca double, i32 16, align 8
12 %ptr = getelementptr double* %array, i32 %b
13 store double %val, double* %ptr, align 8
14 call void @llvm.AMDGPU.barrier.local() noduplicate nounwind
15 %result = load double* %ptr, align 8
16 store double %result, double addrspace(1)* %out, align 8
20 ; SI-LABEL: @indirect_access_v2f64_alloca:
21 ; SI: BUFFER_STORE_DWORDX4
22 define void @v2f64_alloca(<2 x double> addrspace(1)* noalias %out, <2 x double> addrspace(1)* noalias %in, i32 %b) nounwind {
23 %val = load <2 x double> addrspace(1)* %in, align 16
24 %array = alloca <2 x double>, i32 16, align 16
25 %ptr = getelementptr <2 x double>* %array, i32 %b
26 store <2 x double> %val, <2 x double>* %ptr, align 16
27 call void @llvm.AMDGPU.barrier.local() noduplicate nounwind
28 %result = load <2 x double>* %ptr, align 16
29 store <2 x double> %result, <2 x double> addrspace(1)* %out, align 16