28 lines · plain
1; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s | FileCheck -check-prefix=GCN %s2 3@L = external local_unnamed_addr addrspace(3) global [9 x double], align 164@Ldisp = external local_unnamed_addr addrspace(3) global [96 x double], align 165 6; Stores are reordered during loads merge. This case used to assert while7; scanning for a paired instruction because it used to expect paired one8; to follow a base one.9 10; GCN-LABEL: {{^}}out_of_order_merge:11; GCN-COUNT2: ds_read2_b6412; GCN-COUNT3: ds_write_b6413define amdgpu_kernel void @out_of_order_merge() {14entry:15 %gep2 = getelementptr inbounds [96 x double], ptr addrspace(3) @Ldisp, i32 0, i32 116 %tmp12 = load <2 x double>, ptr addrspace(3) getelementptr inbounds ([9 x double], ptr addrspace(3) @L, i32 0, i32 1), align 817 %tmp14 = extractelement <2 x double> %tmp12, i32 018 %tmp15 = extractelement <2 x double> %tmp12, i32 119 %add50.i = fadd double %tmp14, %tmp1520 store double %add50.i, ptr addrspace(3) @Ldisp, align 821 %tmp16 = load double, ptr addrspace(3) getelementptr inbounds ([9 x double], ptr addrspace(3) @L, i32 1, i32 0), align 822 store double %tmp16, ptr addrspace(3) %gep2, align 823 %tmp17 = load <2 x double>, ptr addrspace(3) getelementptr inbounds ([9 x double], ptr addrspace(3) @L, i32 2, i32 1), align 824 %tmp19 = extractelement <2 x double> %tmp17, i32 125 store double %tmp19, ptr addrspace(3) poison, align 826 ret void27}28