74 lines · plain
1; RUN: llc -mtriple=amdgcn -mcpu=tahiti < %s | FileCheck -check-prefixes=CHECK,NOHSA %s2; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=kaveri < %s | FileCheck -check-prefixes=CHECK,HSA %s3; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=fiji < %s | FileCheck -check-prefixes=CHECK,HSA %s4 5; RUN: llc -global-isel -mtriple=amdgcn -mcpu=tahiti < %s | FileCheck -check-prefixes=CHECK,NOHSA %s6; RUN: llc -global-isel -mtriple=amdgcn--amdhsa -mcpu=kaveri < %s | FileCheck -check-prefixes=CHECK,HSA %s7; RUN: llc -global-isel -mtriple=amdgcn--amdhsa -mcpu=fiji < %s | FileCheck -check-prefixes=CHECK,HSA %s8 9@lds0 = addrspace(3) global [512 x float] poison, align 410@lds1 = addrspace(3) global [256 x float] poison, align 411 12@large = addrspace(3) global [4096 x i32] poison, align 413 14; CHECK-LABEL: {{^}}groupstaticsize_test0:15; NOHSA: v_mov_b32_e32 v{{[0-9]+}}, llvm.amdgcn.groupstaticsize@abs32@lo16; HSA: v_mov_b32_e32 v{{[0-9]+}}, 0x800{{$}}17define amdgpu_kernel void @groupstaticsize_test0(ptr addrspace(1) %out, ptr addrspace(1) %lds_size) #0 {18 %tid.x = tail call i32 @llvm.amdgcn.workitem.id.x() #119 %idx.0 = add nsw i32 %tid.x, 6420 %static_lds_size = call i32 @llvm.amdgcn.groupstaticsize() #121 store i32 %static_lds_size, ptr addrspace(1) %lds_size, align 422 %arrayidx0 = getelementptr inbounds [512 x float], ptr addrspace(3) @lds0, i32 0, i32 %idx.023 %val0 = load float, ptr addrspace(3) %arrayidx0, align 424 store float %val0, ptr addrspace(1) %out, align 425 26 ret void27}28 29; CHECK-LABEL: {{^}}groupstaticsize_test1:30; NOHSA: v_mov_b32_e32 v{{[0-9]+}}, llvm.amdgcn.groupstaticsize@abs32@lo31; HSA: v_mov_b32_e32 v{{[0-9]+}}, 0xc00{{$}}32define amdgpu_kernel void @groupstaticsize_test1(ptr addrspace(1) %out, i32 %cond, ptr addrspace(1) %lds_size) {33entry:34 %static_lds_size = call i32 @llvm.amdgcn.groupstaticsize() #135 store i32 %static_lds_size, ptr addrspace(1) %lds_size, align 436 %tid.x = tail call i32 @llvm.amdgcn.workitem.id.x() #137 %idx.0 = add nsw i32 %tid.x, 6438 %tmp = icmp eq i32 %cond, 039 br i1 %tmp, label %if, label %else40 41if: ; preds = %entry42 %arrayidx0 = getelementptr inbounds [512 x float], ptr addrspace(3) @lds0, i32 0, i32 %idx.043 %val0 = load float, ptr addrspace(3) %arrayidx0, align 444 store float %val0, ptr addrspace(1) %out, align 445 br label %endif46 47else: ; preds = %entry48 %arrayidx1 = getelementptr inbounds [256 x float], ptr addrspace(3) @lds1, i32 0, i32 %idx.049 %val1 = load float, ptr addrspace(3) %arrayidx1, align 450 store float %val1, ptr addrspace(1) %out, align 451 br label %endif52 53endif: ; preds = %else, %if54 ret void55}56 57; Exceeds 16-bit simm limit of s_movk_i3258; CHECK-LABEL: {{^}}large_groupstaticsize:59; NOHSA: v_mov_b32_e32 v{{[0-9]+}}, llvm.amdgcn.groupstaticsize@abs32@lo60; HSA: v_mov_b32_e32 [[REG:v[0-9]+]], 0x4000{{$}}61define amdgpu_kernel void @large_groupstaticsize(ptr addrspace(1) %size, i32 %idx) #0 {62 %gep = getelementptr inbounds [4096 x i32], ptr addrspace(3) @large, i32 0, i32 %idx63 store volatile i32 0, ptr addrspace(3) %gep64 %static_lds_size = call i32 @llvm.amdgcn.groupstaticsize()65 store i32 %static_lds_size, ptr addrspace(1) %size66 ret void67}68 69declare i32 @llvm.amdgcn.groupstaticsize() #170declare i32 @llvm.amdgcn.workitem.id.x() #171 72attributes #0 = { nounwind }73attributes #1 = { nounwind readnone }74