31 lines · plain
1; RUN: llc -mtriple=amdgcn -asm-verbose < %s | FileCheck -check-prefix=SI %s2; RUN: llc -mtriple=amdgcn-unknown-amdhsa -asm-verbose -mattr=-flat-for-global < %s | FileCheck -check-prefix=SI %s3 4declare i32 @llvm.amdgcn.mbcnt.lo(i32, i32) #05declare i32 @llvm.amdgcn.mbcnt.hi(i32, i32) #06 7; SI-LABEL: {{^}}foo:8; SI: .section .AMDGPU.csdata9; SI: ; Kernel info:10; SI: ; TotalNumSgprs: {{[0-9]+}}11; SI: ; NumVgprs: {{[0-9]+}}12define amdgpu_kernel void @foo(ptr addrspace(1) noalias %out, ptr addrspace(1) %abase, ptr addrspace(1) %bbase) nounwind {13 %mbcnt.lo = call i32 @llvm.amdgcn.mbcnt.lo(i32 -1, i32 0);14 %tid = call i32 @llvm.amdgcn.mbcnt.hi(i32 -1, i32 %mbcnt.lo)15 %aptr = getelementptr i32, ptr addrspace(1) %abase, i32 %tid16 %bptr = getelementptr i32, ptr addrspace(1) %bbase, i32 %tid17 %outptr = getelementptr i32, ptr addrspace(1) %out, i32 %tid18 %a = load i32, ptr addrspace(1) %aptr, align 419 %b = load i32, ptr addrspace(1) %bptr, align 420 %result = add i32 %a, %b21 store i32 %result, ptr addrspace(1) %outptr, align 422 ret void23}24 25; SI-LABEL: {{^}}one_vgpr_used:26; SI: NumVgprs: 127define amdgpu_kernel void @one_vgpr_used(ptr addrspace(1) %out, i32 %x) nounwind {28 store i32 %x, ptr addrspace(1) %out, align 429 ret void30}31