1; RUN: opt -S -amdgpu-lower-intrinsics %s | FileCheck -check-prefix=OPT %s 2 3declare void @llvm.memcpy.p1i8.p1i8.i64(i8 addrspace(1)* nocapture, i8 addrspace(1)* nocapture readonly, i64, i1) #1 4declare void @llvm.memcpy.p1i8.p3i8.i32(i8 addrspace(1)* nocapture, i8 addrspace(3)* nocapture readonly, i32, i1) #1 5 6declare void @llvm.memmove.p1i8.p1i8.i64(i8 addrspace(1)* nocapture, i8 addrspace(1)* nocapture readonly, i64, i1) #1 7declare void @llvm.memset.p1i8.i64(i8 addrspace(1)* nocapture, i8, i64, i1) #1 8 9; Test the upper bound for sizes to leave 10; OPT-LABEL: @max_size_small_static_memcpy_caller0( 11; OPT: call void @llvm.memcpy.p1i8.p1i8.i64(i8 addrspace(1)* %dst, i8 addrspace(1)* %src, i64 1024, i1 false) 12define amdgpu_kernel void @max_size_small_static_memcpy_caller0(i8 addrspace(1)* %dst, i8 addrspace(1)* %src) #0 { 13 call void @llvm.memcpy.p1i8.p1i8.i64(i8 addrspace(1)* %dst, i8 addrspace(1)* %src, i64 1024, i1 false) 14 ret void 15} 16 17; Smallest static size which will be expanded 18; OPT-LABEL: @min_size_large_static_memcpy_caller0( 19; OPT-NOT: call 20; OPT: br label %load-store-loop 21; OPT: [[T1:%[0-9]+]] = getelementptr inbounds i8, i8 addrspace(1)* %src, i64 %loop-index 22; OPT-NEXT: [[T2:%[0-9]+]] = load i8, i8 addrspace(1)* [[T1]] 23; OPT-NEXT: [[T3:%[0-9]+]] = getelementptr inbounds i8, i8 addrspace(1)* %dst, i64 %loop-index 24; OPT-NEXT: store i8 [[T2]], i8 addrspace(1)* [[T3]] 25; OPT-NEXT: [[T4:%[0-9]+]] = add i64 %loop-index, 1 26; OPT-NEXT: [[T5:%[0-9]+]] = icmp ult i64 [[T4]], 1025 27; OPT-NEXT: br i1 [[T5]], label %load-store-loop, label %memcpy-split 28define amdgpu_kernel void @min_size_large_static_memcpy_caller0(i8 addrspace(1)* %dst, i8 addrspace(1)* %src) #0 { 29 call void @llvm.memcpy.p1i8.p1i8.i64(i8 addrspace(1)* %dst, i8 addrspace(1)* %src, i64 1025, i1 false) 30 ret void 31} 32 33; OPT-LABEL: @max_size_small_static_memmove_caller0( 34; OPT: call void @llvm.memmove.p1i8.p1i8.i64(i8 addrspace(1)* %dst, i8 addrspace(1)* %src, i64 1024, i1 false) 35define amdgpu_kernel void @max_size_small_static_memmove_caller0(i8 addrspace(1)* %dst, i8 addrspace(1)* %src) #0 { 36 call void @llvm.memmove.p1i8.p1i8.i64(i8 addrspace(1)* %dst, i8 addrspace(1)* %src, i64 1024, i1 false) 37 ret void 38} 39 40; OPT-LABEL: @min_size_large_static_memmove_caller0( 41; OPT-NOT: call 42; OPT: getelementptr 43; OPT-NEXT: load i8 44; OPT: getelementptr 45; OPT-NEXT: store i8 46define amdgpu_kernel void @min_size_large_static_memmove_caller0(i8 addrspace(1)* %dst, i8 addrspace(1)* %src) #0 { 47 call void @llvm.memmove.p1i8.p1i8.i64(i8 addrspace(1)* %dst, i8 addrspace(1)* %src, i64 1025, i1 false) 48 ret void 49} 50 51; OPT-LABEL: @max_size_small_static_memset_caller0( 52; OPT: call void @llvm.memset.p1i8.i64(i8 addrspace(1)* %dst, i8 %val, i64 1024, i1 false) 53define amdgpu_kernel void @max_size_small_static_memset_caller0(i8 addrspace(1)* %dst, i8 %val) #0 { 54 call void @llvm.memset.p1i8.i64(i8 addrspace(1)* %dst, i8 %val, i64 1024, i1 false) 55 ret void 56} 57 58; OPT-LABEL: @min_size_large_static_memset_caller0( 59; OPT-NOT: call 60; OPT: getelementptr 61; OPT: store i8 62define amdgpu_kernel void @min_size_large_static_memset_caller0(i8 addrspace(1)* %dst, i8 %val) #0 { 63 call void @llvm.memset.p1i8.i64(i8 addrspace(1)* %dst, i8 %val, i64 1025, i1 false) 64 ret void 65} 66 67; OPT-LABEL: @variable_memcpy_caller0( 68; OPT-NOT: call 69; OPT: phi 70define amdgpu_kernel void @variable_memcpy_caller0(i8 addrspace(1)* %dst, i8 addrspace(1)* %src, i64 %n) #0 { 71 call void @llvm.memcpy.p1i8.p1i8.i64(i8 addrspace(1)* %dst, i8 addrspace(1)* %src, i64 %n, i1 false) 72 ret void 73} 74 75; OPT-LABEL: @variable_memcpy_caller1( 76; OPT-NOT: call 77; OPT: phi 78define amdgpu_kernel void @variable_memcpy_caller1(i8 addrspace(1)* %dst, i8 addrspace(1)* %src, i64 %n) #0 { 79 call void @llvm.memcpy.p1i8.p1i8.i64(i8 addrspace(1)* %dst, i8 addrspace(1)* %src, i64 %n, i1 false) 80 ret void 81} 82 83; OPT-LABEL: @memcpy_multi_use_one_function( 84; OPT-NOT: call 85; OPT: phi 86; OPT-NOT: call 87; OPT: phi 88; OPT-NOT: call 89define amdgpu_kernel void @memcpy_multi_use_one_function(i8 addrspace(1)* %dst0, i8 addrspace(1)* %dst1, i8 addrspace(1)* %src, i64 %n, i64 %m) #0 { 90 call void @llvm.memcpy.p1i8.p1i8.i64(i8 addrspace(1)* %dst0, i8 addrspace(1)* %src, i64 %n, i1 false) 91 call void @llvm.memcpy.p1i8.p1i8.i64(i8 addrspace(1)* %dst1, i8 addrspace(1)* %src, i64 %m, i1 false) 92 ret void 93} 94 95; OPT-LABEL: @memcpy_alt_type( 96; OPT: phi 97; OPT: getelementptr inbounds i8, i8 addrspace(3)* 98; OPT: load i8, i8 addrspace(3)* 99; OPT: getelementptr inbounds i8, i8 addrspace(1)* 100; OPT: store i8 101define amdgpu_kernel void @memcpy_alt_type(i8 addrspace(1)* %dst, i8 addrspace(3)* %src, i32 %n) #0 { 102 call void @llvm.memcpy.p1i8.p3i8.i32(i8 addrspace(1)* %dst, i8 addrspace(3)* %src, i32 %n, i1 false) 103 ret void 104} 105 106; One of the uses in the function should be expanded, the other left alone. 107; OPT-LABEL: @memcpy_multi_use_one_function_keep_small( 108; OPT: getelementptr inbounds i8, i8 addrspace(1)* 109; OPT: load i8, i8 addrspace(1)* 110; OPT: getelementptr inbounds i8, i8 addrspace(1)* 111; OPT: store i8 112 113; OPT: call void @llvm.memcpy.p1i8.p1i8.i64(i8 addrspace(1)* %dst1, i8 addrspace(1)* %src, i64 102, i1 false) 114define amdgpu_kernel void @memcpy_multi_use_one_function_keep_small(i8 addrspace(1)* %dst0, i8 addrspace(1)* %dst1, i8 addrspace(1)* %src, i64 %n) #0 { 115 call void @llvm.memcpy.p1i8.p1i8.i64(i8 addrspace(1)* %dst0, i8 addrspace(1)* %src, i64 %n, i1 false) 116 call void @llvm.memcpy.p1i8.p1i8.i64(i8 addrspace(1)* %dst1, i8 addrspace(1)* %src, i64 102, i1 false) 117 ret void 118} 119 120attributes #0 = { nounwind } 121attributes #1 = { argmemonly nounwind } 122