-
Notifications
You must be signed in to change notification settings - Fork 15.5k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Oneof's Message field builder does nothing when its sub-field changes #16
Comments
We had a discussion about field lifetime issue just recently and here is assumption that I think users should make when accessing sub-fields of a protobuf message: After getting a reference/pointer to a field of a protobuf message, any subsequent mutation method calls to the message may invalidate this reference/pointer and accessing this reference/pointer is an undefined behavior. This not only applies to the case you mentioned, but also applies to other cases and other languages. For example, step 3 in the following scenario will result in undefined behavior:
It might be better to throw an exception to let the user aware of the problem, but given the implementation overhead to address this, I would prefer leaving it as-is. |
Collections' modCount like? :) Understood! David |
Resolve compilation errors if compiled with strict warnings.
…um value is detected. This should never happen, so I don't think it matters much exactly what kind of exception we throw. We could even arguably return null, but this option saves a lot of space while still preserving some error checking. See https://godbolt.org/z/jKhcKs3x1 for code gen. This generates much tighter bytecode and ARM assembly than alternatives. As this code is generated many times over, small wins in code size here can reduce icache pressure, APK size, and OAT size. This java code: ```java Object uoe() { throw new UnsupportedOperationException(); } Object npe2() { throw null; } ``` Generates this dex code: ``` .method uoe()Ljava/lang/Object; new-instance v0, Ljava/lang/UnsupportedOperationException; invoke-direct {v0}, Ljava/lang/UnsupportedOperationException;-><init>()V throw v0 .end method .method npe2()Ljava/lang/Object; const/4 v0, 0x0 throw v0 .end method ``` Which generates this OAT code: ``` java.lang.Object SomeProto.uoe() [84 bytes] 0x000081c0 sub x16, sp, #0x2000 (8192) 0x000081c4 ldr wzr, [x16] StackMap[0] native_pc=0x41c8, dex_pc=0x0, register_mask=0x0, stack_mask=0b 0x000081c8 str x0, [sp, #-48]! 0x000081cc str x22, [sp, #24] 0x000081d0 stp x23, lr, [sp, #32] 0x000081d4 ldr x21, [x21] StackMap[1] native_pc=0x41d8, dex_pc=0x0, register_mask=0x2, stack_mask=0b 0x000081d8 mov x22, x1 0x000081dc adrp x0, #+0x4000 (addr 0x0000c000) 0x000081e0 ldr w0, [x0, #4] 0x000081e4 ldr lr, [tr, #464] ; pAllocObjectInitialized 0x000081e8 blr lr StackMap[2] native_pc=0x41ec, dex_pc=0x0, register_mask=0x400000, stack_mask=0b 0x000081ec dmb ishst 0x000081f0 mov x1, x0 0x000081f4 mov x23, x1 0x000081f8 adrp x0, #+0x4000 (addr 0x0000c000) 0x000081fc ldr w0, [x0, #12] 0x00008200 ldr lr, [x0, #24] 0x00008204 blr lr StackMap[3] native_pc=0x4208, dex_pc=0x2, register_mask=0xc00000, stack_mask=0b 0x00008208 mov x0, x23 0x0000820c ldr lr, [tr, #1264] ; pDeliverException 0x00008210 blr lr StackMap[4] native_pc=0x4214, dex_pc=0x5, register_mask=0xc00000, stack_mask=0b java.lang.Object SomeProto.npe2() [36 bytes] 0x000080d0 sub x16, sp, #0x2000 (8192) 0x000080d4 ldr wzr, [x16] StackMap[0] native_pc=0x40d8, dex_pc=0x0, register_mask=0x0, stack_mask=0b 0x000080d8 str x0, [sp, #-32]! 0x000080dc stp x22, lr, [sp, #16] 0x000080e0 ldr x21, [x21] StackMap[1] native_pc=0x40e4, dex_pc=0x0, register_mask=0x2, stack_mask=0b 0x000080e4 mov x22, x1 0x000080e8 mov w0, #0x0 0x000080ec ldr lr, [tr, #1264] ; pDeliverException 0x000080f0 blr lr StackMap[2] native_pc=0x40f4, dex_pc=0x1, register_mask=0x400000, stack_mask=0b ``` This saves 84-36 = 48 bytes of OAT per method. PiperOrigin-RevId: 684258075
…um value is detected. This should never happen, so I don't think it matters much exactly what kind of exception we throw. We could even arguably return null, but this option saves a lot of space while still preserving some error checking. See https://godbolt.org/z/jKhcKs3x1 for code gen. This generates much tighter bytecode and ARM assembly than alternatives. As this code is generated many times over, small wins in code size here can reduce icache pressure, APK size, and OAT size. This java code: ```java Object uoe() { throw new UnsupportedOperationException(); } Object npe2() { throw null; } ``` Generates this dex code: ``` .method uoe()Ljava/lang/Object; new-instance v0, Ljava/lang/UnsupportedOperationException; invoke-direct {v0}, Ljava/lang/UnsupportedOperationException;-><init>()V throw v0 .end method .method npe2()Ljava/lang/Object; const/4 v0, 0x0 throw v0 .end method ``` Which generates this OAT code: ``` java.lang.Object SomeProto.uoe() [84 bytes] 0x000081c0 sub x16, sp, #0x2000 (8192) 0x000081c4 ldr wzr, [x16] StackMap[0] native_pc=0x41c8, dex_pc=0x0, register_mask=0x0, stack_mask=0b 0x000081c8 str x0, [sp, #-48]! 0x000081cc str x22, [sp, #24] 0x000081d0 stp x23, lr, [sp, #32] 0x000081d4 ldr x21, [x21] StackMap[1] native_pc=0x41d8, dex_pc=0x0, register_mask=0x2, stack_mask=0b 0x000081d8 mov x22, x1 0x000081dc adrp x0, #+0x4000 (addr 0x0000c000) 0x000081e0 ldr w0, [x0, #4] 0x000081e4 ldr lr, [tr, #464] ; pAllocObjectInitialized 0x000081e8 blr lr StackMap[2] native_pc=0x41ec, dex_pc=0x0, register_mask=0x400000, stack_mask=0b 0x000081ec dmb ishst 0x000081f0 mov x1, x0 0x000081f4 mov x23, x1 0x000081f8 adrp x0, #+0x4000 (addr 0x0000c000) 0x000081fc ldr w0, [x0, #12] 0x00008200 ldr lr, [x0, #24] 0x00008204 blr lr StackMap[3] native_pc=0x4208, dex_pc=0x2, register_mask=0xc00000, stack_mask=0b 0x00008208 mov x0, x23 0x0000820c ldr lr, [tr, #1264] ; pDeliverException 0x00008210 blr lr StackMap[4] native_pc=0x4214, dex_pc=0x5, register_mask=0xc00000, stack_mask=0b java.lang.Object SomeProto.npe2() [36 bytes] 0x000080d0 sub x16, sp, #0x2000 (8192) 0x000080d4 ldr wzr, [x16] StackMap[0] native_pc=0x40d8, dex_pc=0x0, register_mask=0x0, stack_mask=0b 0x000080d8 str x0, [sp, #-32]! 0x000080dc stp x22, lr, [sp, #16] 0x000080e0 ldr x21, [x21] StackMap[1] native_pc=0x40e4, dex_pc=0x0, register_mask=0x2, stack_mask=0b 0x000080e4 mov x22, x1 0x000080e8 mov w0, #0x0 0x000080ec ldr lr, [tr, #1264] ; pDeliverException 0x000080f0 blr lr StackMap[2] native_pc=0x40f4, dex_pc=0x1, register_mask=0x400000, stack_mask=0b ``` This saves 84-36 = 48 bytes of OAT per method. PiperOrigin-RevId: 684258075
…um value is detected. This should never happen, so I don't think it matters much exactly what kind of exception we throw. We could even arguably return null, but this option saves a lot of space while still preserving some error checking. See https://godbolt.org/z/jKhcKs3x1 for code gen. This generates much tighter bytecode and ARM assembly than alternatives. As this code is generated many times over, small wins in code size here can reduce icache pressure, APK size, and OAT size. This java code: ```java Object uoe() { throw new UnsupportedOperationException(); } Object npe2() { throw null; } ``` Generates this dex code: ``` .method uoe()Ljava/lang/Object; new-instance v0, Ljava/lang/UnsupportedOperationException; invoke-direct {v0}, Ljava/lang/UnsupportedOperationException;-><init>()V throw v0 .end method .method npe2()Ljava/lang/Object; const/4 v0, 0x0 throw v0 .end method ``` Which generates this OAT code: ``` java.lang.Object SomeProto.uoe() [84 bytes] 0x000081c0 sub x16, sp, #0x2000 (8192) 0x000081c4 ldr wzr, [x16] StackMap[0] native_pc=0x41c8, dex_pc=0x0, register_mask=0x0, stack_mask=0b 0x000081c8 str x0, [sp, #-48]! 0x000081cc str x22, [sp, #24] 0x000081d0 stp x23, lr, [sp, #32] 0x000081d4 ldr x21, [x21] StackMap[1] native_pc=0x41d8, dex_pc=0x0, register_mask=0x2, stack_mask=0b 0x000081d8 mov x22, x1 0x000081dc adrp x0, #+0x4000 (addr 0x0000c000) 0x000081e0 ldr w0, [x0, #4] 0x000081e4 ldr lr, [tr, #464] ; pAllocObjectInitialized 0x000081e8 blr lr StackMap[2] native_pc=0x41ec, dex_pc=0x0, register_mask=0x400000, stack_mask=0b 0x000081ec dmb ishst 0x000081f0 mov x1, x0 0x000081f4 mov x23, x1 0x000081f8 adrp x0, #+0x4000 (addr 0x0000c000) 0x000081fc ldr w0, [x0, #12] 0x00008200 ldr lr, [x0, #24] 0x00008204 blr lr StackMap[3] native_pc=0x4208, dex_pc=0x2, register_mask=0xc00000, stack_mask=0b 0x00008208 mov x0, x23 0x0000820c ldr lr, [tr, #1264] ; pDeliverException 0x00008210 blr lr StackMap[4] native_pc=0x4214, dex_pc=0x5, register_mask=0xc00000, stack_mask=0b java.lang.Object SomeProto.npe2() [36 bytes] 0x000080d0 sub x16, sp, #0x2000 (8192) 0x000080d4 ldr wzr, [x16] StackMap[0] native_pc=0x40d8, dex_pc=0x0, register_mask=0x0, stack_mask=0b 0x000080d8 str x0, [sp, #-32]! 0x000080dc stp x22, lr, [sp, #16] 0x000080e0 ldr x21, [x21] StackMap[1] native_pc=0x40e4, dex_pc=0x0, register_mask=0x2, stack_mask=0b 0x000080e4 mov x22, x1 0x000080e8 mov w0, #0x0 0x000080ec ldr lr, [tr, #1264] ; pDeliverException 0x000080f0 blr lr StackMap[2] native_pc=0x40f4, dex_pc=0x1, register_mask=0x400000, stack_mask=0b ``` This saves 84-36 = 48 bytes of OAT per method. PiperOrigin-RevId: 684258075
…um value is detected. This should never happen, so I don't think it matters much exactly what kind of exception we throw. We could even arguably return null, but this option saves a lot of space while still preserving some error checking. See https://godbolt.org/z/jKhcKs3x1 for code gen. This generates much tighter bytecode and ARM assembly than alternatives. As this code is generated many times over, small wins in code size here can reduce icache pressure, APK size, and OAT size. This java code: ```java Object uoe() { throw new UnsupportedOperationException(); } Object npe2() { throw null; } ``` Generates this dex code: ``` .method uoe()Ljava/lang/Object; new-instance v0, Ljava/lang/UnsupportedOperationException; invoke-direct {v0}, Ljava/lang/UnsupportedOperationException;-><init>()V throw v0 .end method .method npe2()Ljava/lang/Object; const/4 v0, 0x0 throw v0 .end method ``` Which generates this OAT code: ``` java.lang.Object SomeProto.uoe() [84 bytes] 0x000081c0 sub x16, sp, #0x2000 (8192) 0x000081c4 ldr wzr, [x16] StackMap[0] native_pc=0x41c8, dex_pc=0x0, register_mask=0x0, stack_mask=0b 0x000081c8 str x0, [sp, #-48]! 0x000081cc str x22, [sp, #24] 0x000081d0 stp x23, lr, [sp, #32] 0x000081d4 ldr x21, [x21] StackMap[1] native_pc=0x41d8, dex_pc=0x0, register_mask=0x2, stack_mask=0b 0x000081d8 mov x22, x1 0x000081dc adrp x0, #+0x4000 (addr 0x0000c000) 0x000081e0 ldr w0, [x0, #4] 0x000081e4 ldr lr, [tr, #464] ; pAllocObjectInitialized 0x000081e8 blr lr StackMap[2] native_pc=0x41ec, dex_pc=0x0, register_mask=0x400000, stack_mask=0b 0x000081ec dmb ishst 0x000081f0 mov x1, x0 0x000081f4 mov x23, x1 0x000081f8 adrp x0, #+0x4000 (addr 0x0000c000) 0x000081fc ldr w0, [x0, #12] 0x00008200 ldr lr, [x0, #24] 0x00008204 blr lr StackMap[3] native_pc=0x4208, dex_pc=0x2, register_mask=0xc00000, stack_mask=0b 0x00008208 mov x0, x23 0x0000820c ldr lr, [tr, #1264] ; pDeliverException 0x00008210 blr lr StackMap[4] native_pc=0x4214, dex_pc=0x5, register_mask=0xc00000, stack_mask=0b java.lang.Object SomeProto.npe2() [36 bytes] 0x000080d0 sub x16, sp, #0x2000 (8192) 0x000080d4 ldr wzr, [x16] StackMap[0] native_pc=0x40d8, dex_pc=0x0, register_mask=0x0, stack_mask=0b 0x000080d8 str x0, [sp, #-32]! 0x000080dc stp x22, lr, [sp, #16] 0x000080e0 ldr x21, [x21] StackMap[1] native_pc=0x40e4, dex_pc=0x0, register_mask=0x2, stack_mask=0b 0x000080e4 mov x22, x1 0x000080e8 mov w0, #0x0 0x000080ec ldr lr, [tr, #1264] ; pDeliverException 0x000080f0 blr lr StackMap[2] native_pc=0x40f4, dex_pc=0x1, register_mask=0x400000, stack_mask=0b ``` This saves 84-36 = 48 bytes of OAT per method. PiperOrigin-RevId: 684620833
Scenario:
Expected:
There are 2 alternatives:
Actual:
The field builder allows to mutate it with no warnings, and the oneof case is not changed back to this field!
Proposed Remedy:
Add extra oneofFieldChanged(Internal.EnumLite | FieldDescriptor) method to GeneratedMessage.BuilderParent interface, and notify parent of any oneof field builder's mutations.
When notified, either throw an exception if oneof is set to different case, or change the oneof's case to this field's case.
The good news is that it just clarifies the existing behavior and doesn't change any public interfaces and the existing techniques. Not implementing this fix will cause numerous, difficult to trace, logical problems with oneof in such a common scenario.
In addition, due to the wrong bundle number bug, you have to issue minor update anyway; so this would be the opportune time for the proposed solution.
Cheers,
David
PS: Here is the full JUnit 4 test case for the scenario as-is:
import static org.hamcrest.MatcherAssert.assertThat;
import static org.hamcrest.Matchers.equalTo;
import static org.hamcrest.Matchers.is;
import org.junit.Test;
import protobuf_unittest.UnittestProto.TestAllTypes;
import protobuf_unittest.UnittestProto.TestAllTypes.Builder;
import protobuf_unittest.UnittestProto.TestAllTypes.OneofFieldCase;
public class OneofPropertyChangeTest {
@test
public void testOneOfPropertyChangeGeneratedBuilder() {
final Builder builder = TestAllTypes.newBuilder().setOneofString("hello");
assertThat(builder.getOneofFieldCase(), is(equalTo(OneofFieldCase.ONEOF_STRING)));
}
}
The text was updated successfully, but these errors were encountered: