1 ; RUN: opt -mtriple=amdgcn-- -loop-unroll -simplifycfg -sroa %s -S -o - | FileCheck %s 2 ; RUN: opt -mtriple=r600-- -loop-unroll -simplifycfg -sroa %s -S -o - | FileCheck %s 3 4 5 ; This test contains a simple loop that initializes an array declared in 6 ; private memory. We want to make sure these kinds of loops are always 7 ; unrolled, because private memory is slow. 8 9 ; CHECK-LABEL: @private_memory 10 ; CHECK-NOT: alloca 11 ; CHECK: store i32 5, i32 addrspace(1)* %out 12 define amdgpu_kernel void @private_memory(i32 addrspace(1)* %out) { 13 entry: 14 %0 = alloca [32 x i32] 15 br label %loop.header 16 17 loop.header: 18 %counter = phi i32 [0, %entry], [%inc, %loop.inc] 19 br label %loop.body 20 21 loop.body: 22 %ptr = getelementptr [32 x i32], [32 x i32]* %0, i32 0, i32 %counter 23 store i32 %counter, i32* %ptr 24 br label %loop.inc 25 26 loop.inc: 27 %inc = add i32 %counter, 1 28 %1 = icmp sge i32 %counter, 32 29 br i1 %1, label %exit, label %loop.header 30 31 exit: 32 %2 = getelementptr [32 x i32], [32 x i32]* %0, i32 0, i32 5 33 %3 = load i32, i32* %2 34 store i32 %3, i32 addrspace(1)* %out 35 ret void 36 } 37 38 ; Check that loop is unrolled for local memory references 39 40 ; CHECK-LABEL: @local_memory 41 ; CHECK: getelementptr i32, i32 addrspace(1)* %out, i32 128 42 ; CHECK-NEXT: store 43 ; CHECK-NEXT: ret 44 define amdgpu_kernel void @local_memory(i32 addrspace(1)* %out, i32 addrspace(3)* %lds) { 45 entry: 46 br label %loop.header 47 48 loop.header: 49 %counter = phi i32 [0, %entry], [%inc, %loop.inc] 50 br label %loop.body 51 52 loop.body: 53 %ptr_lds = getelementptr i32, i32 addrspace(3)* %lds, i32 %counter 54 %val = load i32, i32 addrspace(3)* %ptr_lds 55 %ptr_out = getelementptr i32, i32 addrspace(1)* %out, i32 %counter 56 store i32 %val, i32 addrspace(1)* %ptr_out 57 br label %loop.inc 58 59 loop.inc: 60 %inc = add i32 %counter, 1 61 %cond = icmp sge i32 %counter, 128 62 br i1 %cond, label %exit, label %loop.header 63 64 exit: 65 ret void 66 } 67 68 ; Check that a loop with if inside completely unrolled to eliminate phi and if 69 70 ; CHECK-LABEL: @unroll_for_if 71 ; CHECK: entry: 72 ; CHECK-NEXT: getelementptr 73 ; CHECK-NEXT: store 74 ; CHECK-NEXT: getelementptr 75 ; CHECK-NEXT: store 76 ; CHECK-NOT: br 77 define amdgpu_kernel void @unroll_for_if(i32* %a) { 78 entry: 79 br label %for.body 80 81 for.body: ; preds = %entry, %for.inc 82 %i1 = phi i32 [ 0, %entry ], [ %inc, %for.inc ] 83 %and = and i32 %i1, 1 84 %tobool = icmp eq i32 %and, 0 85 br i1 %tobool, label %for.inc, label %if.then 86 87 if.then: ; preds = %for.body 88 %0 = sext i32 %i1 to i64 89 %arrayidx = getelementptr inbounds i32, i32* %a, i64 %0 90 store i32 0, i32* %arrayidx, align 4 91 br label %for.inc 92 93 for.inc: ; preds = %for.body, %if.then 94 %inc = add nuw nsw i32 %i1, 1 95 %cmp = icmp ult i32 %inc, 48 96 br i1 %cmp, label %for.body, label %for.end 97 98 for.end: ; preds = %for.cond 99 ret void 100 } 101
