RFR: 8258010: Debug build failure with clang-10 due to -Wdeprecated-copy and -Wimplicit-int-float-conversion [v2]

[Kim Barrett]

> On Jan 3, 2021, at 11:33 PM, Hao Sun <github.com+16932759+shqking at openjdk.java.net> wrote:
> 
> On Mon, 4 Jan 2021 01:18:47 GMT, Hao Sun <github.com+16932759+shqking at openjdk.org> wrote:
> 
>>>> _Mailing list message from [Kim Barrett](mailto:kim.barrett at oracle.com) on [build-dev](mailto:build-dev at openjdk.java.net):_
>> 
>> It appears that either clang is different from gcc for -felide-constructors
>> (which seems unlikely), or clang (clang-10) is doing the deprecation warning
>> at a different point from gcc (quite plausible). That is, clang could be
>> checking for the deprecated case before eliding the call, while gcc is
>> checking for the deprecated case after copy elision has been applied.
> 
> Thanks for your reply.
> I checked the source code of clang-10 and gcc-9 and found that:
> 
> 1) for clang-10,
> 'Wdeprecated-copy' is implemented at the 'Sema' module of clang. See https://github.com/llvm/llvm-project/blob/release/10.x/clang/lib/Sema/SemaDeclCXX.cpp#L13585
> 
> Flag 'felide-constructors' would enable 'felide_constructors' and flag 'fno-elide-constructors' would enables 'fno_elide_constructors'. (See https://github.com/llvm/llvm-project/blob/release/10.x/clang/include/clang/Driver/Options.td).
> Then 'ElideConstructors' will be set (See https://github.com/llvm/llvm-project/blob/release/10.x/clang/lib/Frontend/CompilerInvocation.cpp#L2863)
> Finally, constructors might be elided in several spots in 'CodeGen' module. 
> See: 
> https://github.com/llvm/llvm-project/blob/release/10.x/clang/lib/CodeGen/CGStmt.cpp#L1094
> https://github.com/llvm/llvm-project/blob/release/10.x/clang/lib/CodeGen/CGExprCXX.cpp#L611
> https://github.com/llvm/llvm-project/blob/release/10.x/clang/lib/CodeGen/CGDecl.cpp#L1405
> 
> As far as I know, 'Sema' module is conducted to check semantics errors before 'CodeGen' module.
> That verified your conjecture, i.e. 'clang could be checking for the derepcated case before eliding the call'.
> 
> 2) for gcc-9,
> 'felide-constructors' and 'Wdeprecated-copy' are implemented in a number of spots in gcc. I currently didn't figure out their execution order clearly.
> 
> But in one of the use points at function build_over_call(), 'flag_elide_constructors' (See https://github.com/gcc-mirror/gcc/blob/releases/gcc-9/gcc/cp/call.c#L8538) is handled **before** 'warn_deprecated_copy' (See https://github.com/gcc-mirror/gcc/blob/releases/gcc-9/gcc/cp/call.c#L8608 and https://github.com/gcc-mirror/gcc/blob/releases/gcc-9/gcc/cp/call.c#L8679)
> 
> Hope that my finding is helpful. Thanks.

Yes, that would explain the different warning behaviors.  Thanks for digging through that.