139 lines · plain
1; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=gfx900 -o - %s | FileCheck %s2 3@lds0 = addrspace(3) global [512 x float] poison4@lds1 = addrspace(3) global [256 x float] poison5@lds2 = addrspace(3) global [4096 x float] poison6@lds3 = addrspace(3) global [67 x i8] poison7 8@dynamic_shared0 = external addrspace(3) global [0 x float]9@dynamic_shared1 = external addrspace(3) global [0 x double]10@dynamic_shared2 = external addrspace(3) global [0 x double], align 411@dynamic_shared3 = external addrspace(3) global [0 x double], align 1612 13; CHECK-LABEL: {{^}}dynamic_shared_array_0:14; CHECK: v_add_u32_e32 v{{[0-9]+}}, 0x800, v{{[0-9]+}}15define amdgpu_kernel void @dynamic_shared_array_0(ptr addrspace(1) %out) {16 %tid.x = tail call i32 @llvm.amdgcn.workitem.id.x()17 %arrayidx0 = getelementptr inbounds [512 x float], ptr addrspace(3) @lds0, i32 0, i32 %tid.x18 %val0 = load float, ptr addrspace(3) %arrayidx0, align 419 %arrayidx1 = getelementptr inbounds [0 x float], ptr addrspace(3) @dynamic_shared0, i32 0, i32 %tid.x20 store float %val0, ptr addrspace(3) %arrayidx1, align 421 ret void22}23 24; CHECK-LABEL: {{^}}dynamic_shared_array_1:25; CHECK: s_movk_i32 [[DYNLDS:s[0-9]+]], 0xc0026; CHECK: v_lshl_add_u32 {{v[0-9]+}}, {{v[0-9]+}}, 2, [[DYNLDS]]27define amdgpu_kernel void @dynamic_shared_array_1(ptr addrspace(1) %out, i32 %cond) {28entry:29 %tid.x = tail call i32 @llvm.amdgcn.workitem.id.x()30 %idx.0 = add nsw i32 %tid.x, 6431 %tmp = icmp eq i32 %cond, 032 br i1 %tmp, label %if, label %else33 34if: ; preds = %entry35 %arrayidx0 = getelementptr inbounds [512 x float], ptr addrspace(3) @lds0, i32 0, i32 %idx.036 %val0 = load float, ptr addrspace(3) %arrayidx0, align 437 br label %endif38 39else: ; preds = %entry40 %arrayidx1 = getelementptr inbounds [256 x float], ptr addrspace(3) @lds1, i32 0, i32 %idx.041 %val1 = load float, ptr addrspace(3) %arrayidx1, align 442 br label %endif43 44endif: ; preds = %else, %if45 %val = phi float [ %val0, %if ], [ %val1, %else ]46 %arrayidx = getelementptr inbounds [0 x float], ptr addrspace(3) @dynamic_shared0, i32 0, i32 %tid.x47 store float %val, ptr addrspace(3) %arrayidx, align 448 ret void49}50 51; CHECK-LABEL: {{^}}dynamic_shared_array_2:52; CHECK: s_movk_i32 [[DYNLDS:s[0-9]+]], 0x400053; CHECK: v_lshl_add_u32 {{v[0-9]+}}, {{v[0-9]+}}, 2, [[DYNLDS]]54define amdgpu_kernel void @dynamic_shared_array_2(i32 %idx) {55 %tid.x = tail call i32 @llvm.amdgcn.workitem.id.x()56 %vidx = add i32 %tid.x, %idx57 %arrayidx0 = getelementptr inbounds [4096 x float], ptr addrspace(3) @lds2, i32 0, i32 %vidx58 %val0 = load float, ptr addrspace(3) %arrayidx0, align 459 %arrayidx1 = getelementptr inbounds [0 x float], ptr addrspace(3) @dynamic_shared0, i32 0, i32 %tid.x60 store float %val0, ptr addrspace(3) %arrayidx1, align 461 ret void62}63 64; The offset to the dynamic shared memory array should be aligned on the type65; specified.66; CHECK-LABEL: {{^}}dynamic_shared_array_3:67; CHECK: s_movk_i32 [[DYNLDS:s[0-9]+]], 0x4468; CHECK: v_lshl_add_u32 {{v[0-9]+}}, {{v[0-9]+}}, 2, [[DYNLDS]]69define amdgpu_kernel void @dynamic_shared_array_3(i32 %idx) {70 %tid.x = tail call i32 @llvm.amdgcn.workitem.id.x()71 %vidx = add i32 %tid.x, %idx72 %arrayidx0 = getelementptr inbounds [67 x i8], ptr addrspace(3) @lds3, i32 0, i32 %vidx73 %val0 = load i8, ptr addrspace(3) %arrayidx0, align 474 %val1 = uitofp i8 %val0 to float75 %arrayidx1 = getelementptr inbounds [0 x float], ptr addrspace(3) @dynamic_shared0, i32 0, i32 %tid.x76 store float %val1, ptr addrspace(3) %arrayidx1, align 477 ret void78}79 80; The offset to the dynamic shared memory array should be aligned on the81; maximal one.82; CHECK-LABEL: {{^}}dynamic_shared_array_4:83; CHECK: s_movk_i32 [[DYNLDS:s[0-9]+]], 0x4884; CHECK-DAG: v_lshl_add_u32 {{v[0-9]+}}, {{v[0-9]+}}, 2, [[DYNLDS]]85; CHECK-DAG: v_lshl_add_u32 {{v[0-9]+}}, {{v[0-9]+}}, 3, [[DYNLDS]]86define amdgpu_kernel void @dynamic_shared_array_4(i32 %idx) {87 %tid.x = tail call i32 @llvm.amdgcn.workitem.id.x()88 %vidx = add i32 %tid.x, %idx89 %arrayidx0 = getelementptr inbounds [67 x i8], ptr addrspace(3) @lds3, i32 0, i32 %vidx90 %val0 = load i8, ptr addrspace(3) %arrayidx0, align 491 %val1 = uitofp i8 %val0 to float92 %val2 = uitofp i8 %val0 to double93 %arrayidx1 = getelementptr inbounds [0 x float], ptr addrspace(3) @dynamic_shared0, i32 0, i32 %tid.x94 store float %val1, ptr addrspace(3) %arrayidx1, align 495 %arrayidx2 = getelementptr inbounds [0 x double], ptr addrspace(3) @dynamic_shared1, i32 0, i32 %tid.x96 store double %val2, ptr addrspace(3) %arrayidx2, align 497 ret void98}99 100; Honor the explicit alignment from the specified variable.101; CHECK-LABEL: {{^}}dynamic_shared_array_5:102; CHECK: s_movk_i32 [[DYNLDS:s[0-9]+]], 0x44103; CHECK-DAG: v_lshl_add_u32 {{v[0-9]+}}, {{v[0-9]+}}, 2, [[DYNLDS]]104; CHECK-DAG: v_lshl_add_u32 {{v[0-9]+}}, {{v[0-9]+}}, 3, [[DYNLDS]]105define amdgpu_kernel void @dynamic_shared_array_5(i32 %idx) {106 %tid.x = tail call i32 @llvm.amdgcn.workitem.id.x()107 %vidx = add i32 %tid.x, %idx108 %arrayidx0 = getelementptr inbounds [67 x i8], ptr addrspace(3) @lds3, i32 0, i32 %vidx109 %val0 = load i8, ptr addrspace(3) %arrayidx0, align 4110 %val1 = uitofp i8 %val0 to float111 %val2 = uitofp i8 %val0 to double112 %arrayidx1 = getelementptr inbounds [0 x float], ptr addrspace(3) @dynamic_shared0, i32 0, i32 %tid.x113 store float %val1, ptr addrspace(3) %arrayidx1, align 4114 %arrayidx2 = getelementptr inbounds [0 x double], ptr addrspace(3) @dynamic_shared2, i32 0, i32 %tid.x115 store double %val2, ptr addrspace(3) %arrayidx2, align 4116 ret void117}118 119; Honor the explicit alignment from the specified variable.120; CHECK-LABEL: {{^}}dynamic_shared_array_6:121; CHECK: s_movk_i32 [[DYNLDS:s[0-9]+]], 0x50122; CHECK-DAG: v_lshl_add_u32 {{v[0-9]+}}, {{v[0-9]+}}, 2, [[DYNLDS]]123; CHECK-DAG: v_lshl_add_u32 {{v[0-9]+}}, {{v[0-9]+}}, 3, [[DYNLDS]]124define amdgpu_kernel void @dynamic_shared_array_6(i32 %idx) {125 %tid.x = tail call i32 @llvm.amdgcn.workitem.id.x()126 %vidx = add i32 %tid.x, %idx127 %arrayidx0 = getelementptr inbounds [67 x i8], ptr addrspace(3) @lds3, i32 0, i32 %vidx128 %val0 = load i8, ptr addrspace(3) %arrayidx0, align 4129 %val1 = uitofp i8 %val0 to float130 %val2 = uitofp i8 %val0 to double131 %arrayidx1 = getelementptr inbounds [0 x float], ptr addrspace(3) @dynamic_shared0, i32 0, i32 %tid.x132 store float %val1, ptr addrspace(3) %arrayidx1, align 4133 %arrayidx2 = getelementptr inbounds [0 x double], ptr addrspace(3) @dynamic_shared3, i32 0, i32 %tid.x134 store double %val2, ptr addrspace(3) %arrayidx2, align 4135 ret void136}137 138declare i32 @llvm.amdgcn.workitem.id.x()139