143 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 52; RUN: llc < %s -mtriple=nvptx64 | FileCheck %s3; RUN: %if ptxas %{ llc < %s -mtriple=nvptx64 | %ptxas-verify %}4 5target triple = "nvptx64-nvidia-cuda"6 7%struct.T = type { i64, <2 x i32>, <4 x i32> }8 9declare void @test_call(%struct.T)10 11define void @test_store_param_undef() {12; CHECK-LABEL: test_store_param_undef(13; CHECK: {14; CHECK-EMPTY:15; CHECK-EMPTY:16; CHECK-NEXT: // %bb.0:17; CHECK-NEXT: { // callseq 0, 018; CHECK-NEXT: .param .align 16 .b8 param0[32];19; CHECK-NEXT: call.uni test_call, (param0);20; CHECK-NEXT: } // callseq 021; CHECK-NEXT: ret;22 call void @test_call(%struct.T undef)23 ret void24}25 26define void @test_store_param_def(i64 %param0, i32 %param1) {27; CHECK-LABEL: test_store_param_def(28; CHECK: {29; CHECK-NEXT: .reg .b32 %r<6>;30; CHECK-NEXT: .reg .b64 %rd<2>;31; CHECK-EMPTY:32; CHECK-NEXT: // %bb.0:33; CHECK-NEXT: ld.param.b64 %rd1, [test_store_param_def_param_0];34; CHECK-NEXT: ld.param.b32 %r1, [test_store_param_def_param_1];35; CHECK-NEXT: { // callseq 1, 036; CHECK-NEXT: .param .align 16 .b8 param0[32];37; CHECK-NEXT: st.param.v4.b32 [param0+16], {%r2, %r1, %r3, %r4};38; CHECK-NEXT: st.param.v2.b32 [param0+8], {%r5, %r1};39; CHECK-NEXT: st.param.b64 [param0], %rd1;40; CHECK-NEXT: call.uni test_call, (param0);41; CHECK-NEXT: } // callseq 142; CHECK-NEXT: ret;43 %V2 = insertelement <2 x i32> undef, i32 %param1, i32 144 %V4 = insertelement <4 x i32> undef, i32 %param1, i32 145 %S0 = insertvalue %struct.T undef, i64 %param0, 046 %S1 = insertvalue %struct.T %S0, <2 x i32> %V2, 147 %S2 = insertvalue %struct.T %S1, <4 x i32> %V4, 248 call void @test_call(%struct.T %S2)49 ret void50}51 52define void @test_store_undef(ptr %out) {53; CHECK-LABEL: test_store_undef(54; CHECK: {55; CHECK-EMPTY:56; CHECK-EMPTY:57; CHECK-NEXT: // %bb.0:58; CHECK-NEXT: ret;59 store %struct.T undef, ptr %out60 ret void61}62 63define void @test_store_def(i64 %param0, i32 %param1, ptr %out) {64; CHECK-LABEL: test_store_def(65; CHECK: {66; CHECK-NEXT: .reg .b32 %r<6>;67; CHECK-NEXT: .reg .b64 %rd<3>;68; CHECK-EMPTY:69; CHECK-NEXT: // %bb.0:70; CHECK-NEXT: ld.param.b64 %rd1, [test_store_def_param_0];71; CHECK-NEXT: ld.param.b32 %r1, [test_store_def_param_1];72; CHECK-NEXT: ld.param.b64 %rd2, [test_store_def_param_2];73; CHECK-NEXT: st.v4.b32 [%rd2+16], {%r2, %r1, %r3, %r4};74; CHECK-NEXT: st.v2.b32 [%rd2+8], {%r5, %r1};75; CHECK-NEXT: st.b64 [%rd2], %rd1;76; CHECK-NEXT: ret;77 %V2 = insertelement <2 x i32> undef, i32 %param1, i32 178 %V4 = insertelement <4 x i32> undef, i32 %param1, i32 179 %S0 = insertvalue %struct.T undef, i64 %param0, 080 %S1 = insertvalue %struct.T %S0, <2 x i32> %V2, 181 %S2 = insertvalue %struct.T %S1, <4 x i32> %V4, 282 store %struct.T %S2, ptr %out83 ret void84}85 86define void @test_store_volatile_undef(ptr %out, <8 x i32> %vec) {87; CHECK-LABEL: test_store_volatile_undef(88; CHECK: {89; CHECK-NEXT: .reg .b32 %r<23>;90; CHECK-NEXT: .reg .b64 %rd<5>;91; CHECK-EMPTY:92; CHECK-NEXT: // %bb.0:93; CHECK-NEXT: ld.param.b64 %rd1, [test_store_volatile_undef_param_0];94; CHECK-NEXT: st.volatile.v4.b32 [%rd1+16], {%r1, %r2, %r3, %r4};95; CHECK-NEXT: st.volatile.v2.b32 [%rd1+8], {%r5, %r6};96; CHECK-NEXT: st.volatile.b64 [%rd1], %rd2;97; CHECK-NEXT: ld.param.v4.b32 {%r7, %r8, %r9, %r10}, [test_store_volatile_undef_param_1];98; CHECK-NEXT: ld.param.v4.b32 {%r11, %r12, %r13, %r14}, [test_store_volatile_undef_param_1+16];99; CHECK-NEXT: st.volatile.v4.b32 [%rd3], {%r11, %r12, %r13, %r14};100; CHECK-NEXT: st.volatile.v4.b32 [%rd4], {%r7, %r8, %r9, %r10};101; CHECK-NEXT: st.volatile.v4.b32 [%rd1+16], {%r15, %r16, %r17, %r18};102; CHECK-NEXT: st.volatile.v4.b32 [%rd1], {%r19, %r20, %r21, %r22};103; CHECK-NEXT: ret;104 store volatile %struct.T undef, ptr %out105 store volatile <8 x i32> %vec, ptr undef106 store volatile <8 x i32> undef, ptr %out107 ret void108}109 110define void @test_store_volatile_of_poison(ptr %out) {111; CHECK-LABEL: test_store_volatile_of_poison(112; CHECK: {113; CHECK-NEXT: .reg .b32 %r<7>;114; CHECK-NEXT: .reg .b64 %rd<3>;115; CHECK-EMPTY:116; CHECK-NEXT: // %bb.0:117; CHECK-NEXT: ld.param.b64 %rd1, [test_store_volatile_of_poison_param_0];118; CHECK-NEXT: st.volatile.v4.b32 [%rd1+16], {%r1, %r2, %r3, %r4};119; CHECK-NEXT: st.volatile.v2.b32 [%rd1+8], {%r5, %r6};120; CHECK-NEXT: st.volatile.b64 [%rd1], %rd2;121; CHECK-NEXT: ret;122 store volatile %struct.T poison, ptr %out123 ret void124}125 126define void @test_store_volatile_to_poison(%struct.T %param) {127; CHECK-LABEL: test_store_volatile_to_poison(128; CHECK: {129; CHECK-NEXT: .reg .b32 %r<7>;130; CHECK-NEXT: .reg .b64 %rd<5>;131; CHECK-EMPTY:132; CHECK-NEXT: // %bb.0:133; CHECK-NEXT: ld.param.b64 %rd1, [test_store_volatile_to_poison_param_0];134; CHECK-NEXT: ld.param.v2.b32 {%r1, %r2}, [test_store_volatile_to_poison_param_0+8];135; CHECK-NEXT: ld.param.v4.b32 {%r3, %r4, %r5, %r6}, [test_store_volatile_to_poison_param_0+16];136; CHECK-NEXT: st.volatile.v4.b32 [%rd2], {%r3, %r4, %r5, %r6};137; CHECK-NEXT: st.volatile.v2.b32 [%rd3], {%r1, %r2};138; CHECK-NEXT: st.volatile.b64 [%rd4], %rd1;139; CHECK-NEXT: ret;140 store volatile %struct.T %param, ptr poison141 ret void142}143