; NOTE: Assertions have been autogenerated by utils/update_test_checks.py ; RUN: opt -dse -S < %s | FileCheck %s declare void @llvm.lifetime.start.p0i8(i64 immarg, i8* nocapture) declare void @llvm.lifetime.end.p0i8(i64 immarg, i8* nocapture) declare void @unknown() declare void @f(i8*) declare void @f2(i8*, i8*) declare i8* @f3(i8*, i8*) ; Basic case for DSEing a trivially dead writing call define void @test_dead() { ; CHECK-LABEL: @test_dead( ; CHECK-NEXT: ret void ; %a = alloca i32, align 4 %bitcast = bitcast i32* %a to i8* call void @f(i8* writeonly nocapture %bitcast) argmemonly nounwind willreturn ret void } ; Add in canonical lifetime intrinsics define void @test_lifetime() { ; CHECK-LABEL: @test_lifetime( ; CHECK-NEXT: [[A:%.*]] = alloca i32, align 4 ; CHECK-NEXT: [[BITCAST:%.*]] = bitcast i32* [[A]] to i8* ; CHECK-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[BITCAST]]) ; CHECK-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[BITCAST]]) ; CHECK-NEXT: ret void ; %a = alloca i32, align 4 %bitcast = bitcast i32* %a to i8* call void @llvm.lifetime.start.p0i8(i64 4, i8* %bitcast) call void @f(i8* writeonly nocapture %bitcast) argmemonly nounwind willreturn call void @llvm.lifetime.end.p0i8(i64 4, i8* %bitcast) ret void } ; Add some unknown calls just to point out that this is use based, not ; instruction order sensitive define void @test_lifetime2() { ; CHECK-LABEL: @test_lifetime2( ; CHECK-NEXT: [[A:%.*]] = alloca i32, align 4 ; CHECK-NEXT: [[BITCAST:%.*]] = bitcast i32* [[A]] to i8* ; CHECK-NEXT: call void @llvm.lifetime.start.p0i8(i64 4, i8* [[BITCAST]]) ; CHECK-NEXT: call void @unknown() ; CHECK-NEXT: call void @unknown() ; CHECK-NEXT: call void @llvm.lifetime.end.p0i8(i64 4, i8* [[BITCAST]]) ; CHECK-NEXT: ret void ; %a = alloca i32, align 4 %bitcast = bitcast i32* %a to i8* call void @llvm.lifetime.start.p0i8(i64 4, i8* %bitcast) call void @unknown() call void @f(i8* writeonly nocapture %bitcast) argmemonly nounwind willreturn call void @unknown() call void @llvm.lifetime.end.p0i8(i64 4, i8* %bitcast) ret void } ; As long as the result is unused, we can even remove reads of the alloca ; itself since the write will be dropped. define void @test_dead_readwrite() { ; CHECK-LABEL: @test_dead_readwrite( ; CHECK-NEXT: ret void ; %a = alloca i32, align 4 %bitcast = bitcast i32* %a to i8* call void @f(i8* nocapture %bitcast) argmemonly nounwind willreturn ret void } define i32 @test_neg_read_after() { ; CHECK-LABEL: @test_neg_read_after( ; CHECK-NEXT: [[A:%.*]] = alloca i32, align 4 ; CHECK-NEXT: [[BITCAST:%.*]] = bitcast i32* [[A]] to i8* ; CHECK-NEXT: call void @f(i8* nocapture writeonly [[BITCAST]]) #[[ATTR1:[0-9]+]] ; CHECK-NEXT: [[RES:%.*]] = load i32, i32* [[A]], align 4 ; CHECK-NEXT: ret i32 [[RES]] ; %a = alloca i32, align 4 %bitcast = bitcast i32* %a to i8* call void @f(i8* writeonly nocapture %bitcast) argmemonly nounwind willreturn %res = load i32, i32* %a ret i32 %res } define void @test_neg_infinite_loop() { ; CHECK-LABEL: @test_neg_infinite_loop( ; CHECK-NEXT: [[A:%.*]] = alloca i32, align 4 ; CHECK-NEXT: [[BITCAST:%.*]] = bitcast i32* [[A]] to i8* ; CHECK-NEXT: call void @f(i8* nocapture writeonly [[BITCAST]]) #[[ATTR2:[0-9]+]] ; CHECK-NEXT: ret void ; %a = alloca i32, align 4 %bitcast = bitcast i32* %a to i8* call void @f(i8* writeonly nocapture %bitcast) argmemonly nounwind ret void } define void @test_neg_throw() { ; CHECK-LABEL: @test_neg_throw( ; CHECK-NEXT: [[A:%.*]] = alloca i32, align 4 ; CHECK-NEXT: [[BITCAST:%.*]] = bitcast i32* [[A]] to i8* ; CHECK-NEXT: call void @f(i8* nocapture writeonly [[BITCAST]]) #[[ATTR3:[0-9]+]] ; CHECK-NEXT: ret void ; %a = alloca i32, align 4 %bitcast = bitcast i32* %a to i8* call void @f(i8* writeonly nocapture %bitcast) argmemonly willreturn ret void } define void @test_neg_extra_write() { ; CHECK-LABEL: @test_neg_extra_write( ; CHECK-NEXT: [[A:%.*]] = alloca i32, align 4 ; CHECK-NEXT: [[BITCAST:%.*]] = bitcast i32* [[A]] to i8* ; CHECK-NEXT: call void @f(i8* nocapture writeonly [[BITCAST]]) #[[ATTR4:[0-9]+]] ; CHECK-NEXT: ret void ; %a = alloca i32, align 4 %bitcast = bitcast i32* %a to i8* call void @f(i8* writeonly nocapture %bitcast) nounwind willreturn ret void } ; In this case, we can't remove a1 because we need to preserve the write to ; a2, and if we leave the call around, we need memory to pass to the first arg. define void @test_neg_unmodeled_write() { ; CHECK-LABEL: @test_neg_unmodeled_write( ; CHECK-NEXT: [[A:%.*]] = alloca i32, align 4 ; CHECK-NEXT: [[A2:%.*]] = alloca i32, align 4 ; CHECK-NEXT: [[BITCAST:%.*]] = bitcast i32* [[A]] to i8* ; CHECK-NEXT: [[BITCAST2:%.*]] = bitcast i32* [[A2]] to i8* ; CHECK-NEXT: call void @f2(i8* nocapture writeonly [[BITCAST]], i8* [[BITCAST2]]) #[[ATTR1]] ; CHECK-NEXT: ret void ; %a = alloca i32, align 4 %a2 = alloca i32, align 4 %bitcast = bitcast i32* %a to i8* %bitcast2 = bitcast i32* %a2 to i8* call void @f2(i8* nocapture writeonly %bitcast, i8* %bitcast2) argmemonly nounwind willreturn ret void } define i32 @test_neg_captured_by_call() { ; CHECK-LABEL: @test_neg_captured_by_call( ; CHECK-NEXT: [[A:%.*]] = alloca i32, align 4 ; CHECK-NEXT: [[A2:%.*]] = alloca i8*, align 4 ; CHECK-NEXT: [[BITCAST:%.*]] = bitcast i32* [[A]] to i8* ; CHECK-NEXT: [[BITCAST2:%.*]] = bitcast i8** [[A2]] to i8* ; CHECK-NEXT: call void @f2(i8* writeonly [[BITCAST]], i8* [[BITCAST2]]) #[[ATTR1]] ; CHECK-NEXT: [[A_COPY_CAST:%.*]] = load i8*, i8** [[A2]], align 8 ; CHECK-NEXT: [[A_COPY:%.*]] = bitcast i8* [[A_COPY_CAST]] to i32* ; CHECK-NEXT: [[RES:%.*]] = load i32, i32* [[A_COPY]], align 4 ; CHECK-NEXT: ret i32 [[RES]] ; %a = alloca i32, align 4 %a2 = alloca i8*, align 4 %bitcast = bitcast i32* %a to i8* %bitcast2 = bitcast i8** %a2 to i8* call void @f2(i8* writeonly %bitcast, i8* %bitcast2) argmemonly nounwind willreturn %a_copy_cast = load i8*, i8** %a2 %a_copy = bitcast i8* %a_copy_cast to i32* %res = load i32, i32* %a_copy ret i32 %res } define i32 @test_neg_captured_before() { ; CHECK-LABEL: @test_neg_captured_before( ; CHECK-NEXT: [[A:%.*]] = alloca i32, align 4 ; CHECK-NEXT: [[A2:%.*]] = alloca i8*, align 4 ; CHECK-NEXT: [[BITCAST:%.*]] = bitcast i32* [[A]] to i8* ; CHECK-NEXT: [[BITCAST2:%.*]] = bitcast i8** [[A2]] to i8* ; CHECK-NEXT: store i8* [[BITCAST]], i8** [[A2]], align 8 ; CHECK-NEXT: call void @f(i8* nocapture writeonly [[BITCAST]]) #[[ATTR1]] ; CHECK-NEXT: [[A_COPY_CAST:%.*]] = load i8*, i8** [[A2]], align 8 ; CHECK-NEXT: [[A_COPY:%.*]] = bitcast i8* [[A_COPY_CAST]] to i32* ; CHECK-NEXT: [[RES:%.*]] = load i32, i32* [[A_COPY]], align 4 ; CHECK-NEXT: ret i32 [[RES]] ; %a = alloca i32, align 4 %a2 = alloca i8*, align 4 %bitcast = bitcast i32* %a to i8* %bitcast2 = bitcast i8** %a2 to i8* store i8* %bitcast, i8** %a2 call void @f(i8* writeonly nocapture %bitcast) argmemonly nounwind willreturn %a_copy_cast = load i8*, i8** %a2 %a_copy = bitcast i8* %a_copy_cast to i32* %res = load i32, i32* %a_copy ret i32 %res } ; Callee might be dead, but op bundle has unknown semantics and thus isn't. define void @test_new_op_bundle() { ; CHECK-LABEL: @test_new_op_bundle( ; CHECK-NEXT: [[A:%.*]] = alloca i32, align 4 ; CHECK-NEXT: [[BITCAST:%.*]] = bitcast i32* [[A]] to i8* ; CHECK-NEXT: call void @f(i8* nocapture writeonly [[BITCAST]]) #[[ATTR1]] [ "unknown"(i8* [[BITCAST]]) ] ; CHECK-NEXT: ret void ; %a = alloca i32, align 4 %bitcast = bitcast i32* %a to i8* call void @f(i8* writeonly nocapture %bitcast) argmemonly nounwind willreturn ["unknown" (i8* %bitcast)] ret void } ; Show that reading from unrelated memory is okay define void @test_unreleated_read() { ; CHECK-LABEL: @test_unreleated_read( ; CHECK-NEXT: ret void ; %a = alloca i32, align 4 %a2 = alloca i32, align 4 %bitcast = bitcast i32* %a to i8* %bitcast2 = bitcast i32* %a2 to i8* call void @f2(i8* nocapture writeonly %bitcast, i8* nocapture readonly %bitcast2) argmemonly nounwind willreturn ret void } ; Removing a capture is also okay. The capture can only be in the return value ; (which is unused) or written into the dead out parameter. define void @test_unrelated_capture() { ; CHECK-LABEL: @test_unrelated_capture( ; CHECK-NEXT: ret void ; %a = alloca i32, align 4 %a2 = alloca i32, align 4 %bitcast = bitcast i32* %a to i8* %bitcast2 = bitcast i32* %a2 to i8* call i8* @f3(i8* nocapture writeonly %bitcast, i8* readonly %bitcast2) argmemonly nounwind willreturn ret void } ; Cannot remove call, as %bitcast2 is captured via the return value. define i8 @test_neg_unrelated_capture_used_via_return() { ; CHECK-LABEL: @test_neg_unrelated_capture_used_via_return( ; CHECK-NEXT: [[A:%.*]] = alloca i32, align 4 ; CHECK-NEXT: [[A2:%.*]] = alloca i32, align 4 ; CHECK-NEXT: [[BITCAST:%.*]] = bitcast i32* [[A]] to i8* ; CHECK-NEXT: [[BITCAST2:%.*]] = bitcast i32* [[A2]] to i8* ; CHECK-NEXT: [[CAPTURE:%.*]] = call i8* @f3(i8* nocapture writeonly [[BITCAST]], i8* readonly [[BITCAST2]]) #[[ATTR1]] ; CHECK-NEXT: [[V:%.*]] = load i8, i8* [[CAPTURE]], align 1 ; CHECK-NEXT: ret i8 [[V]] ; %a = alloca i32, align 4 %a2 = alloca i32, align 4 %bitcast = bitcast i32* %a to i8* %bitcast2 = bitcast i32* %a2 to i8* %capture = call i8* @f3(i8* nocapture writeonly %bitcast, i8* readonly %bitcast2) argmemonly nounwind willreturn %v = load i8, i8* %capture ret i8 %v } ; As long as the result is unused, we can even remove reads of the alloca ; itself since the write will be dropped. define void @test_self_read() { ; CHECK-LABEL: @test_self_read( ; CHECK-NEXT: ret void ; %a = alloca i32, align 4 %bitcast = bitcast i32* %a to i8* call void @f2(i8* nocapture writeonly %bitcast, i8* nocapture readonly %bitcast) argmemonly nounwind willreturn ret void } ; We can remove the call because while we don't know the size of the write done ; by the call, we do know the following store writes to the entire contents of ; the alloca. define i32 @test_dse_overwrite() { ; CHECK-LABEL: @test_dse_overwrite( ; CHECK-NEXT: [[A:%.*]] = alloca i32, align 4 ; CHECK-NEXT: store i32 0, i32* [[A]], align 4 ; CHECK-NEXT: [[V:%.*]] = load i32, i32* [[A]], align 4 ; CHECK-NEXT: ret i32 [[V]] ; %a = alloca i32, align 4 %bitcast = bitcast i32* %a to i8* call void @f(i8* writeonly nocapture %bitcast) argmemonly nounwind willreturn store i32 0, i32* %a %v = load i32, i32* %a ret i32 %v } ; Negative case where we can read part of the value written by @f. define i32 @test_neg_dse_partial_overwrite() { ; CHECK-LABEL: @test_neg_dse_partial_overwrite( ; CHECK-NEXT: [[A:%.*]] = alloca i32, align 4 ; CHECK-NEXT: [[BITCAST:%.*]] = bitcast i32* [[A]] to i8* ; CHECK-NEXT: call void @f(i8* nocapture writeonly [[BITCAST]]) #[[ATTR1]] ; CHECK-NEXT: store i8 0, i8* [[BITCAST]], align 1 ; CHECK-NEXT: [[V:%.*]] = load i32, i32* [[A]], align 4 ; CHECK-NEXT: ret i32 [[V]] ; %a = alloca i32, align 4 %bitcast = bitcast i32* %a to i8* call void @f(i8* writeonly nocapture %bitcast) argmemonly nounwind willreturn store i8 0, i8* %bitcast %v = load i32, i32* %a ret i32 %v } ; Negative case where we don't know the size of a, and thus can't use the ; full overwrite reasoning define i32 @test_neg_dse_unsized(i32* %a) { ; CHECK-LABEL: @test_neg_dse_unsized( ; CHECK-NEXT: [[BITCAST:%.*]] = bitcast i32* [[A:%.*]] to i8* ; CHECK-NEXT: call void @f(i8* nocapture writeonly [[BITCAST]]) #[[ATTR1]] ; CHECK-NEXT: store i32 0, i32* [[A]], align 4 ; CHECK-NEXT: [[V:%.*]] = load i32, i32* [[A]], align 4 ; CHECK-NEXT: ret i32 [[V]] ; %bitcast = bitcast i32* %a to i8* call void @f(i8* writeonly nocapture %bitcast) argmemonly nounwind willreturn store i32 0, i32* %a %v = load i32, i32* %a ret i32 %v } @G = global i8 0 ; Same as test_dse_overwrite, but with a non-alloca object. define void @test_dse_non_alloca() { ; CHECK-LABEL: @test_dse_non_alloca( ; CHECK-NEXT: store i8 0, i8* @G, align 1 ; CHECK-NEXT: ret void ; call void @f(i8* writeonly nocapture @G) argmemonly nounwind willreturn store i8 0, i8* @G ret void }