153 lines · plain
1; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=r600 -mcpu=redwood < %s | FileCheck -enable-var-scope -check-prefixes=EG,FUNC %s2 3declare i32 @llvm.r600.read.tidig.x() readnone4 5; FUNC-LABEL: {{^}}s_sub_i32:6define amdgpu_kernel void @s_sub_i32(ptr addrspace(1) %out, i32 %a, i32 %b) {7 %result = sub i32 %a, %b8 store i32 %result, ptr addrspace(1) %out9 ret void10}11 12; FUNC-LABEL: {{^}}s_sub_imm_i32:13define amdgpu_kernel void @s_sub_imm_i32(ptr addrspace(1) %out, i32 %a) {14 %result = sub i32 1234, %a15 store i32 %result, ptr addrspace(1) %out16 ret void17}18 19; FUNC-LABEL: {{^}}test_sub_i32:20; EG: SUB_INT {{\** *}}T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}21define amdgpu_kernel void @test_sub_i32(ptr addrspace(1) %out, ptr addrspace(1) %in) {22 %b_ptr = getelementptr i32, ptr addrspace(1) %in, i32 123 %a = load i32, ptr addrspace(1) %in24 %b = load i32, ptr addrspace(1) %b_ptr25 %result = sub i32 %a, %b26 store i32 %result, ptr addrspace(1) %out27 ret void28}29 30; FUNC-LABEL: {{^}}test_sub_imm_i32:31; EG: SUB_INT32define amdgpu_kernel void @test_sub_imm_i32(ptr addrspace(1) %out, ptr addrspace(1) %in) {33 %a = load i32, ptr addrspace(1) %in34 %result = sub i32 123, %a35 store i32 %result, ptr addrspace(1) %out36 ret void37}38 39; FUNC-LABEL: {{^}}test_sub_v2i32:40; EG: SUB_INT {{\** *}}T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}41; EG: SUB_INT {{\** *}}T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}42define amdgpu_kernel void @test_sub_v2i32(ptr addrspace(1) %out, ptr addrspace(1) %in) {43 %b_ptr = getelementptr <2 x i32>, ptr addrspace(1) %in, i32 144 %a = load <2 x i32>, ptr addrspace(1) %in45 %b = load <2 x i32>, ptr addrspace(1) %b_ptr46 %result = sub <2 x i32> %a, %b47 store <2 x i32> %result, ptr addrspace(1) %out48 ret void49}50 51; FUNC-LABEL: {{^}}test_sub_v4i32:52; EG: SUB_INT {{\** *}}T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}53; EG: SUB_INT {{\** *}}T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}54; EG: SUB_INT {{\** *}}T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}55; EG: SUB_INT {{\** *}}T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}56define amdgpu_kernel void @test_sub_v4i32(ptr addrspace(1) %out, ptr addrspace(1) %in) {57 %b_ptr = getelementptr <4 x i32>, ptr addrspace(1) %in, i32 158 %a = load <4 x i32>, ptr addrspace(1) %in59 %b = load <4 x i32>, ptr addrspace(1) %b_ptr60 %result = sub <4 x i32> %a, %b61 store <4 x i32> %result, ptr addrspace(1) %out62 ret void63}64 65; FUNC-LABEL: {{^}}test_sub_i16:66define amdgpu_kernel void @test_sub_i16(ptr addrspace(1) %out, ptr addrspace(1) %in) {67 %tid = call i32 @llvm.r600.read.tidig.x()68 %gep = getelementptr i16, ptr addrspace(1) %in, i32 %tid69 %b_ptr = getelementptr i16, ptr addrspace(1) %gep, i32 170 %a = load volatile i16, ptr addrspace(1) %gep71 %b = load volatile i16, ptr addrspace(1) %b_ptr72 %result = sub i16 %a, %b73 store i16 %result, ptr addrspace(1) %out74 ret void75}76 77; FUNC-LABEL: {{^}}test_sub_v2i16:78define amdgpu_kernel void @test_sub_v2i16(ptr addrspace(1) %out, ptr addrspace(1) %in) {79 %tid = call i32 @llvm.r600.read.tidig.x()80 %gep = getelementptr <2 x i16>, ptr addrspace(1) %in, i32 %tid81 %b_ptr = getelementptr <2 x i16>, ptr addrspace(1) %gep, i16 182 %a = load <2 x i16>, ptr addrspace(1) %gep83 %b = load <2 x i16>, ptr addrspace(1) %b_ptr84 %result = sub <2 x i16> %a, %b85 store <2 x i16> %result, ptr addrspace(1) %out86 ret void87}88 89; FUNC-LABEL: {{^}}test_sub_v4i16:90define amdgpu_kernel void @test_sub_v4i16(ptr addrspace(1) %out, ptr addrspace(1) %in) {91 %tid = call i32 @llvm.r600.read.tidig.x()92 %gep = getelementptr <4 x i16>, ptr addrspace(1) %in, i32 %tid93 %b_ptr = getelementptr <4 x i16>, ptr addrspace(1) %gep, i16 194 %a = load <4 x i16>, ptr addrspace(1) %gep95 %b = load <4 x i16>, ptr addrspace(1) %b_ptr96 %result = sub <4 x i16> %a, %b97 store <4 x i16> %result, ptr addrspace(1) %out98 ret void99}100 101; FUNC-LABEL: {{^}}s_sub_i64:102; EG: MEM_RAT_CACHELESS STORE_RAW T{{[0-9]+}}.XY103; EG-DAG: SUB_INT {{[* ]*}}104; EG-DAG: SUBB_UINT105; EG-DAG: SUB_INT106; EG-DAG: SUB_INT {{[* ]*}}107define amdgpu_kernel void @s_sub_i64(ptr addrspace(1) noalias %out, i64 %a, i64 %b) nounwind {108 %result = sub i64 %a, %b109 store i64 %result, ptr addrspace(1) %out, align 8110 ret void111}112 113; FUNC-LABEL: {{^}}v_sub_i64:114; EG: MEM_RAT_CACHELESS STORE_RAW T{{[0-9]+}}.XY115; EG-DAG: SUB_INT {{[* ]*}}116; EG-DAG: SUBB_UINT117; EG-DAG: SUB_INT118; EG-DAG: SUB_INT {{[* ]*}}119define amdgpu_kernel void @v_sub_i64(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %inA, ptr addrspace(1) noalias %inB) nounwind {120 %tid = call i32 @llvm.r600.read.tidig.x() readnone121 %a_ptr = getelementptr i64, ptr addrspace(1) %inA, i32 %tid122 %b_ptr = getelementptr i64, ptr addrspace(1) %inB, i32 %tid123 %a = load i64, ptr addrspace(1) %a_ptr124 %b = load i64, ptr addrspace(1) %b_ptr125 %result = sub i64 %a, %b126 store i64 %result, ptr addrspace(1) %out, align 8127 ret void128}129 130; FUNC-LABEL: {{^}}v_test_sub_v2i64:131define amdgpu_kernel void @v_test_sub_v2i64(ptr addrspace(1) %out, ptr addrspace(1) noalias %inA, ptr addrspace(1) noalias %inB) {132 %tid = call i32 @llvm.r600.read.tidig.x() readnone133 %a_ptr = getelementptr <2 x i64>, ptr addrspace(1) %inA, i32 %tid134 %b_ptr = getelementptr <2 x i64>, ptr addrspace(1) %inB, i32 %tid135 %a = load <2 x i64>, ptr addrspace(1) %a_ptr136 %b = load <2 x i64>, ptr addrspace(1) %b_ptr137 %result = sub <2 x i64> %a, %b138 store <2 x i64> %result, ptr addrspace(1) %out139 ret void140}141 142; FUNC-LABEL: {{^}}v_test_sub_v4i64:143define amdgpu_kernel void @v_test_sub_v4i64(ptr addrspace(1) %out, ptr addrspace(1) noalias %inA, ptr addrspace(1) noalias %inB) {144 %tid = call i32 @llvm.r600.read.tidig.x() readnone145 %a_ptr = getelementptr <4 x i64>, ptr addrspace(1) %inA, i32 %tid146 %b_ptr = getelementptr <4 x i64>, ptr addrspace(1) %inB, i32 %tid147 %a = load <4 x i64>, ptr addrspace(1) %a_ptr148 %b = load <4 x i64>, ptr addrspace(1) %b_ptr149 %result = sub <4 x i64> %a, %b150 store <4 x i64> %result, ptr addrspace(1) %out151 ret void152}153