diff options
Diffstat (limited to 'test/Transforms/ObjCARC/nested.ll')
| -rw-r--r-- | test/Transforms/ObjCARC/nested.ll | 141 |
1 files changed, 137 insertions, 4 deletions
diff --git a/test/Transforms/ObjCARC/nested.ll b/test/Transforms/ObjCARC/nested.ll index 9eada8a..a618a21 100644 --- a/test/Transforms/ObjCARC/nested.ll +++ b/test/Transforms/ObjCARC/nested.ll @@ -484,12 +484,14 @@ forcoll.empty: ret void } -; Delete a nested retain+release pair. +; TODO: Delete a nested retain+release pair. +; The optimizer currently can't do this, because of a split loop backedge. +; See test9b for the same testcase without a split backedge. ; CHECK: define void @test9( ; CHECK: call i8* @objc_retain ; CHECK: call i8* @objc_retain -; CHECK-NOT: @objc_retain +; CHECK: call i8* @objc_retain ; CHECK: } define void @test9() nounwind { entry: @@ -551,13 +553,79 @@ forcoll.empty: ret void } -; Delete a nested retain+release pair. +; Like test9, but without a split backedge. This we can optimize. -; CHECK: define void @test10( +; CHECK: define void @test9b( ; CHECK: call i8* @objc_retain ; CHECK: call i8* @objc_retain ; CHECK-NOT: @objc_retain ; CHECK: } +define void @test9b() nounwind { +entry: + %state.ptr = alloca %struct.__objcFastEnumerationState, align 8 + %items.ptr = alloca [16 x i8*], align 8 + %call = call i8* @returner() + %0 = call i8* @objc_retainAutoreleasedReturnValue(i8* %call) nounwind + %call1 = call i8* @returner() + %1 = call i8* @objc_retainAutoreleasedReturnValue(i8* %call1) nounwind + %tmp = bitcast %struct.__objcFastEnumerationState* %state.ptr to i8* + call void @llvm.memset.p0i8.i64(i8* %tmp, i8 0, i64 64, i32 8, i1 false) + %2 = call i8* @objc_retain(i8* %0) nounwind + %tmp3 = load i8** @"\01L_OBJC_SELECTOR_REFERENCES_", align 8 + %call4 = call i64 bitcast (i8* (i8*, i8*, ...)* @objc_msgSend to i64 (i8*, i8*, %struct.__objcFastEnumerationState*, [16 x i8*]*, i64)*)(i8* %2, i8* %tmp3, %struct.__objcFastEnumerationState* %state.ptr, [16 x i8*]* %items.ptr, i64 16) + %iszero = icmp eq i64 %call4, 0 + br i1 %iszero, label %forcoll.empty, label %forcoll.loopinit + +forcoll.loopinit: + %mutationsptr.ptr = getelementptr inbounds %struct.__objcFastEnumerationState* %state.ptr, i64 0, i32 2 + %mutationsptr = load i64** %mutationsptr.ptr, align 8 + %forcoll.initial-mutations = load i64* %mutationsptr, align 8 + br label %forcoll.loopbody.outer + +forcoll.loopbody.outer: + %forcoll.count.ph = phi i64 [ %call4, %forcoll.loopinit ], [ %call7, %forcoll.refetch ] + %tmp9 = icmp ugt i64 %forcoll.count.ph, 1 + %umax = select i1 %tmp9, i64 %forcoll.count.ph, i64 1 + br label %forcoll.loopbody + +forcoll.loopbody: + %forcoll.index = phi i64 [ %phitmp, %forcoll.notmutated ], [ 0, %forcoll.loopbody.outer ] + %mutationsptr5 = load i64** %mutationsptr.ptr, align 8 + %statemutations = load i64* %mutationsptr5, align 8 + %3 = icmp eq i64 %statemutations, %forcoll.initial-mutations + br i1 %3, label %forcoll.notmutated, label %forcoll.mutated + +forcoll.mutated: + call void @objc_enumerationMutation(i8* %2) + br label %forcoll.notmutated + +forcoll.notmutated: + %phitmp = add i64 %forcoll.index, 1 + %exitcond = icmp eq i64 %phitmp, %umax + br i1 %exitcond, label %forcoll.refetch, label %forcoll.loopbody + +forcoll.refetch: + %tmp6 = load i8** @"\01L_OBJC_SELECTOR_REFERENCES_", align 8 + %call7 = call i64 bitcast (i8* (i8*, i8*, ...)* @objc_msgSend to i64 (i8*, i8*, %struct.__objcFastEnumerationState*, [16 x i8*]*, i64)*)(i8* %2, i8* %tmp6, %struct.__objcFastEnumerationState* %state.ptr, [16 x i8*]* %items.ptr, i64 16) + %4 = icmp eq i64 %call7, 0 + br i1 %4, label %forcoll.empty, label %forcoll.loopbody.outer + +forcoll.empty: + call void @objc_release(i8* %2) nounwind + call void @objc_release(i8* %1) nounwind, !clang.imprecise_release !0 + call void @objc_release(i8* %0) nounwind, !clang.imprecise_release !0 + ret void +} + +; TODO: Delete a nested retain+release pair. +; The optimizer currently can't do this, because of a split loop backedge. +; See test10b for the same testcase without a split backedge. + +; CHECK: define void @test10( +; CHECK: call i8* @objc_retain +; CHECK: call i8* @objc_retain +; CHECK: call i8* @objc_retain +; CHECK: } define void @test10() nounwind { entry: %state.ptr = alloca %struct.__objcFastEnumerationState, align 8 @@ -618,3 +686,68 @@ forcoll.empty: call void @objc_release(i8* %0) nounwind, !clang.imprecise_release !0 ret void } + +; Like test10, but without a split backedge. This we can optimize. + +; CHECK: define void @test10b( +; CHECK: call i8* @objc_retain +; CHECK: call i8* @objc_retain +; CHECK-NOT: @objc_retain +; CHECK: } +define void @test10b() nounwind { +entry: + %state.ptr = alloca %struct.__objcFastEnumerationState, align 8 + %items.ptr = alloca [16 x i8*], align 8 + %call = call i8* @returner() + %0 = call i8* @objc_retainAutoreleasedReturnValue(i8* %call) nounwind + %call1 = call i8* @returner() + %1 = call i8* @objc_retainAutoreleasedReturnValue(i8* %call1) nounwind + call void @callee() + %tmp = bitcast %struct.__objcFastEnumerationState* %state.ptr to i8* + call void @llvm.memset.p0i8.i64(i8* %tmp, i8 0, i64 64, i32 8, i1 false) + %2 = call i8* @objc_retain(i8* %0) nounwind + %tmp3 = load i8** @"\01L_OBJC_SELECTOR_REFERENCES_", align 8 + %call4 = call i64 bitcast (i8* (i8*, i8*, ...)* @objc_msgSend to i64 (i8*, i8*, %struct.__objcFastEnumerationState*, [16 x i8*]*, i64)*)(i8* %2, i8* %tmp3, %struct.__objcFastEnumerationState* %state.ptr, [16 x i8*]* %items.ptr, i64 16) + %iszero = icmp eq i64 %call4, 0 + br i1 %iszero, label %forcoll.empty, label %forcoll.loopinit + +forcoll.loopinit: + %mutationsptr.ptr = getelementptr inbounds %struct.__objcFastEnumerationState* %state.ptr, i64 0, i32 2 + %mutationsptr = load i64** %mutationsptr.ptr, align 8 + %forcoll.initial-mutations = load i64* %mutationsptr, align 8 + br label %forcoll.loopbody.outer + +forcoll.loopbody.outer: + %forcoll.count.ph = phi i64 [ %call4, %forcoll.loopinit ], [ %call7, %forcoll.refetch ] + %tmp9 = icmp ugt i64 %forcoll.count.ph, 1 + %umax = select i1 %tmp9, i64 %forcoll.count.ph, i64 1 + br label %forcoll.loopbody + +forcoll.loopbody: + %forcoll.index = phi i64 [ %phitmp, %forcoll.notmutated ], [ 0, %forcoll.loopbody.outer ] + %mutationsptr5 = load i64** %mutationsptr.ptr, align 8 + %statemutations = load i64* %mutationsptr5, align 8 + %3 = icmp eq i64 %statemutations, %forcoll.initial-mutations + br i1 %3, label %forcoll.notmutated, label %forcoll.mutated + +forcoll.mutated: + call void @objc_enumerationMutation(i8* %2) + br label %forcoll.notmutated + +forcoll.notmutated: + %phitmp = add i64 %forcoll.index, 1 + %exitcond = icmp eq i64 %phitmp, %umax + br i1 %exitcond, label %forcoll.refetch, label %forcoll.loopbody + +forcoll.refetch: + %tmp6 = load i8** @"\01L_OBJC_SELECTOR_REFERENCES_", align 8 + %call7 = call i64 bitcast (i8* (i8*, i8*, ...)* @objc_msgSend to i64 (i8*, i8*, %struct.__objcFastEnumerationState*, [16 x i8*]*, i64)*)(i8* %2, i8* %tmp6, %struct.__objcFastEnumerationState* %state.ptr, [16 x i8*]* %items.ptr, i64 16) + %4 = icmp eq i64 %call7, 0 + br i1 %4, label %forcoll.empty, label %forcoll.loopbody.outer + +forcoll.empty: + call void @objc_release(i8* %2) nounwind + call void @objc_release(i8* %1) nounwind, !clang.imprecise_release !0 + call void @objc_release(i8* %0) nounwind, !clang.imprecise_release !0 + ret void +} |
