brintos

brintos / llvm-project-archived public Read only

0
0
Text · 3.9 KiB · 095c202 Raw
68 lines · plain
1// NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py UTC_ARGS: --version 52// REQUIRES: amdgpu-registered-target3// RUN: %clang_cc1 -triple amdgcn-amd-amdhsa -target-cpu gfx950 -emit-llvm -fcuda-is-device -o - %s | FileCheck %s4 5// COM: Most tests are in the OpenCL semastics, this is just a verification for HIP6 7#define __device__ __attribute__((device))8#define __shared__ __attribute__((shared))9 10typedef unsigned int u32;11 12// CHECK-LABEL: define dso_local void @_Z20test_load_to_lds_u32PjS_(13// CHECK-SAME: ptr noundef [[SRC:%.*]], ptr noundef [[DST:%.*]]) #[[ATTR0:[0-9]+]] {14// CHECK-NEXT:  [[ENTRY:.*:]]15// CHECK-NEXT:    [[SRC_ADDR:%.*]] = alloca ptr, align 8, addrspace(5)16// CHECK-NEXT:    [[DST_ADDR:%.*]] = alloca ptr, align 8, addrspace(5)17// CHECK-NEXT:    [[SRC_ADDR_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[SRC_ADDR]] to ptr18// CHECK-NEXT:    [[DST_ADDR_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[DST_ADDR]] to ptr19// CHECK-NEXT:    store ptr [[SRC]], ptr [[SRC_ADDR_ASCAST]], align 820// CHECK-NEXT:    store ptr [[DST]], ptr [[DST_ADDR_ASCAST]], align 821// CHECK-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[SRC_ADDR_ASCAST]], align 822// CHECK-NEXT:    [[TMP1:%.*]] = load ptr, ptr [[DST_ADDR_ASCAST]], align 823// CHECK-NEXT:    [[TMP2:%.*]] = addrspacecast ptr [[TMP1]] to ptr addrspace(3)24// CHECK-NEXT:    call void @llvm.amdgcn.load.to.lds.p0(ptr [[TMP0]], ptr addrspace(3) [[TMP2]], i32 4, i32 0, i32 0)25// CHECK-NEXT:    ret void26//27__device__ void test_load_to_lds_u32(u32* src, __shared__ u32 *dst) {28  __builtin_amdgcn_load_to_lds(src, dst, /*size=*/4, /*offset=*/0, /*aux=*/0);29}30 31// CHECK-LABEL: define dso_local void @_Z30test_load_to_lds_u32_flat_destPjS_(32// CHECK-SAME: ptr noundef [[SRC:%.*]], ptr noundef [[DST:%.*]]) #[[ATTR0]] {33// CHECK-NEXT:  [[ENTRY:.*:]]34// CHECK-NEXT:    [[SRC_ADDR:%.*]] = alloca ptr, align 8, addrspace(5)35// CHECK-NEXT:    [[DST_ADDR:%.*]] = alloca ptr, align 8, addrspace(5)36// CHECK-NEXT:    [[SRC_ADDR_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[SRC_ADDR]] to ptr37// CHECK-NEXT:    [[DST_ADDR_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[DST_ADDR]] to ptr38// CHECK-NEXT:    store ptr [[SRC]], ptr [[SRC_ADDR_ASCAST]], align 839// CHECK-NEXT:    store ptr [[DST]], ptr [[DST_ADDR_ASCAST]], align 840// CHECK-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[SRC_ADDR_ASCAST]], align 841// CHECK-NEXT:    [[TMP1:%.*]] = load ptr, ptr [[DST_ADDR_ASCAST]], align 842// CHECK-NEXT:    [[TMP2:%.*]] = addrspacecast ptr [[TMP1]] to ptr addrspace(3)43// CHECK-NEXT:    call void @llvm.amdgcn.load.to.lds.p0(ptr [[TMP0]], ptr addrspace(3) [[TMP2]], i32 4, i32 0, i32 0)44// CHECK-NEXT:    ret void45//46__device__ void test_load_to_lds_u32_flat_dest(u32* src, u32 *dst) {47  __builtin_amdgcn_load_to_lds(src, dst, /*size=*/4, /*offset=*/0, /*aux=*/0);48}49 50// CHECK-LABEL: define dso_local void @_Z20test_load_to_lds_128PvS_(51// CHECK-SAME: ptr noundef [[SRC:%.*]], ptr noundef [[DST:%.*]]) #[[ATTR0]] {52// CHECK-NEXT:  [[ENTRY:.*:]]53// CHECK-NEXT:    [[SRC_ADDR:%.*]] = alloca ptr, align 8, addrspace(5)54// CHECK-NEXT:    [[DST_ADDR:%.*]] = alloca ptr, align 8, addrspace(5)55// CHECK-NEXT:    [[SRC_ADDR_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[SRC_ADDR]] to ptr56// CHECK-NEXT:    [[DST_ADDR_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[DST_ADDR]] to ptr57// CHECK-NEXT:    store ptr [[SRC]], ptr [[SRC_ADDR_ASCAST]], align 858// CHECK-NEXT:    store ptr [[DST]], ptr [[DST_ADDR_ASCAST]], align 859// CHECK-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[SRC_ADDR_ASCAST]], align 860// CHECK-NEXT:    [[TMP1:%.*]] = load ptr, ptr [[DST_ADDR_ASCAST]], align 861// CHECK-NEXT:    [[TMP2:%.*]] = addrspacecast ptr [[TMP1]] to ptr addrspace(3)62// CHECK-NEXT:    call void @llvm.amdgcn.load.to.lds.p0(ptr [[TMP0]], ptr addrspace(3) [[TMP2]], i32 16, i32 0, i32 0)63// CHECK-NEXT:    ret void64//65__device__ void test_load_to_lds_128(void* src, __shared__ void *dst) {66  __builtin_amdgcn_load_to_lds(src, dst, /*size=*/16, /*offset=*/0, /*aux=*/0);67}68